Radical Empiricism, Empirical Modelling and the nature of knowing
Meurig Beynon
Computer Science, University of Warwick, Coventry CV4 7AL

Abstract
This paper explores connections between Radical Empiricism (RE), a 
philosophic attitude developed by William James at the beginning of the 
20th century, and Empirical Modelling (EM), an approach to computer-
based modelling that has been developed by the author and his 
collaborators over a number of years. It focuses in particular on how both 
RE and EM promote a perspective on the nature of knowing that is 
radically different from that typically invoked in contemporary approaches 
to knowledge representation in computing. This is illustrated in detail with 
reference to the modelling of several scenarios of lift use. Some potential 
implications for knowledge management are briefly reviewed.

This paper considers the potential significance of William James's philosophic attitude of 
'Radical Empiricism (RE)' [13] in relation to contemporary problems of knowledge 
representation in the information sciences. Current trends in computer technology and use 
provide a strong motivation for reviewing RE in this light. For instance, as Gooding 
relates in [11], our understanding of how scientific knowledge relates to interaction with 
the natural world and with our peers is challenged by the development of virtual reality 
environments, and the role that virtual experiments have come to play in science. Such 
considerations have prompted a reappraisal of the fundamental assumptions that inform 
the logicist approach to knowledge representation in AI, and called into question the 
extent to which knowledge is mediated by language rather than engagement with the 
world (cf. Cantwell-Smith [9], Turner [16]). In this connection, the relevance of RE 
stems from the priority it ascribes to 'pure experience', and its contention that (to 
paraphrase William James) the whole of the nature of knowing can be put into 
experiential terms ([13], p.56). The problematic aspect of RE, as identified by Bird in [8], 
is that it is of its essence inarticulate: " ... James's pure experience has to be such that 
nothing can be said about it, if it is to fulfil the role for which it is cast". This distances 
RE both from the mainstream philosophical traditions, and from the received views of 
computer programming as intrinsically bound up with formal languages and logical 
specification.

Empirical Modelling (EM) is an approach to computer-based modelling that has been 
developed by the author and his collaborators at the University of Warwick over several 
years (for background, see [1,2,3,4,5,6,7,15,17,18]). The product of an EM exercise is 
first and foremost to be regarded as a source of experience whose interpretation by the 
modeller is not preconceived, but is to be established in the mind of the modeller through 
an association between experience of the model and experience external to the model. 
Knowledge in such a context has the qualities that James attributes to knowledge in [13]: 
it is a personal awareness on the part of the modeller that one experience stands in a 
particular relation to another. To borrow James's expression, experience of interaction 
with the model is 'an experience that knows another' that can act as a substitute for 
experience external to the model in a definite practical sense ([13], p.61).

This paper explores the extent to which, through model-building using EM, it is possible 
to track James's exposition of the empirical roots of knowledge, with its emphasis on the 
fundamental significance in sense-making of our capacity to experience conjunctive 
relations between things. This exploration touches on many issues topical in modern 
computing that are addressed in James's account of RE, such as the nature of 
consciousness ([13], p.132-3), agency ([13], p.178-80, 185-6) and reality ([13], p.159-
60). In what sense, and to what extent, it is possible to establish meaningful connections 
between James's philosophic attitude and EM is itself potentially a controversial issue. 
The author's justification for proposing such connections stems from his own direct 
experience Ð in particular from the way in which James's discussion of 'pure experience' 
in [13] resonates with the issues involved in a detailed exposition of modelling with 
definitive scripts. In James's terms, the thrust of the exposition will be to make it 
plausible that the experience of EM 'knows' pure experience.

The paper is in two principal sections. Section 1 reviews Empirical Modelling (EM) 
principles and practice. Section 2 discusses William James's philosophic attitude of 
Radical Empiricism and the parallels that may be drawn between James's account of 'pure 
experience' and experience of EM. The paper concludes by identifying significant issues 
in knowledge management for which RE and EM may be seen as particularly relevant.

1. Empirical Modelling
	
Empirical Modelling (EM) is a term that has been introduced to describe principles and 
tools that support an unusual kind of computer-based model-building. The development 
of this approach has been the subject of an extended research programme at the 
University of Warwick that originated with the design of a notation for interactive 
graphics some 20 years ago (see [17]). EM is unusual in that it represents a form of 
modelling oriented towards capturing 'state-as-experienced' and leads to the construction 
of computer-based artefacts that have no preconceived or formally circumscribed 
behaviour. It will be helpful to put this claim in context before presenting concrete 
evidence to illustrate it.

1.1. The nature of EM models

It is a commonplace fundamental notion of computer science that the significant 
semantics of a computer program is captured in the algorithmic behaviour that it 
implements. This notion leads on to the idea that all legitimate computer use is 
necessarily essentially concerned with specifying and implementing abstract behaviours. 
Whilst computer science acknowledges the need to design interfaces through which the 
user can direct or monitor computer behaviours, this is typically seen as beyond the remit 
of the core science of computing. Such a view of core computer science as fundamentally 
concerned with abstract mathematical concepts of computation and behaviour is curious 
in view of key trends in the historical development of computing practice.

The use of the computer Ð or more precisely of computer-related technology Ð to generate 
so-called virtual reality (VR) environments epitomises some of the most significant 
issues. If a VR environment is to have the qualities of a reality such as we experience in 
our everyday life, the character of the interface to the underlying computer model of state 
is a crucial rather a peripheral issue. In a real environment, the user can observe the 
environment in ways that have not been preconceived, and conduct authentic 
measurements and experiments. It may be possible in principle to conceive how such a 
real environment can be implemented if we accept a reductionist view of reality and a 
logicist account of human intelligence. Such a conception is of little interest either to the 
computer practitioner or to the user, whose view of construction and use is guided by 
pragmatic concerns. For instance, we should not necessarily expect to have perfect 
knowledge in order to construct an environment to assist us in the task of 'knowledge 
management'. In any event, an environment in which only knowledge that has been 
previously encoded is recoverable is of limited use.

Viewed from the classical computational perspective, the idea of building computer-
based models that offer the user more than has been consciously encoded by way of 'use 
cases' (cf. [12]) seems paradoxical. To see this from another perspective, it is helpful to 
consider other model-building activities, such as those associated with what Levi-Strauss 
characterises as bricolage [14]. In bricolage, the modeller's concept of the artefact under 
construction develops in conjunction with the artefact itself: the modeller gains feedback 
from experience of the unformed artefact itself, and uses this feedback to guide its further 
development. To the extent that prototyping plays a part in computer program 
development, practical computer programming can be viewed in this light. In practice, 
feedback routinely affects the bricoleur's conception of the developing artefact in a much 
more radical way than prototyping affects software development. The bricoleur is not 
constrained by the pre-established conventions for interpretation that typically frame a 
software product: in contrast, the way in which the artefact is to be interpreted and used is 
in many aspects ill-defined and open to negotiation throughout its construction. EM may 
be seen as offering an approach to model-building with the computer that has these key 
characteristics in common with bricolage.

The classical view of computation is complementary to a classical view of knowledge. 
The concept of building programs that are optimised to serve a specific narrow function, 
and of encoding information in formal data structures, promotes a prosaic view of what 
knowledge and knowledge representation entail. Knowledge is seen as something to be 
possessed that can be expressed and recorded as a proposition, as in "I know the 
telephone number of staff member X". This view is appropriate in its proper context, as is 
implied by the use of the expressions 'perfect knowledge' and 'only knowledge that has 
been previously encoded' above. It makes sense to speak of recording the telephone 
numbers of all university staff in a directory, and Ð in the absence of any further context Ð 
it would be absurd to search this directory for those who are blue-eyed and can play the 
bassoon. On the other hand, it is evident that developing a VR environment involves 
much more than encoding abstract knowledge about a real-world domain. A key issue is 
how knowledge about the domain is reflected in the interactive experience that the 
environment offers to the user.

EM differs from classical computer-based model-building in its relation to both 
computation and knowledge. The distinction between an EM model and a computer 
program is analogous to the 'real-world' distinction between an open environment and a 
circumscribed procedure. To be familiar with an EM model is analogous to being familiar 
with a city; to be familiar with a computer program is analogous to knowing how to use 
the underground to get from one station to another. It is helpful to think of the EM model 
as something organic that grows and changes over time: there is no sense in which being 
familiar with a city is a limited notion Ð we explore and observe more of a city over time; 
the city changes; we change in our response to the city; we change the city. There is no 
reasonable sense in which familiarity with a city can be comprehensively preconceived; 
in any moment, we can at most testify to what is familiar in the particular aspect of the 
city of which we have immediate and direct experience. Though I may tell you with 
conviction that the cathedral is out of sight but lies just around the corner, I appeal to my 
memory and to my faith in the permanence of place, and know what only the act of 
taking us to the cathedral can confirm. An EM model is more general than a computer 
program in much the same way that 'knowing how to use the underground' sits within the 
broader framework of 'being familiar with the city'. The analogy also helps to illustrate 
the ontological distinction between the model and the program. If we take the 
permanence and reliability of a city environment and the generic and routine nature of 
our observation of underground transport for granted, then 'knowing how to use the 
underground' is a skill that can be viewed as independent of any particular city, that can 
be exercised without engaging with the total experience of 'becoming familiar with the 
city'. It is possible to imagine how a robot can be programmed with the stimulus-response 
patterns needed to use the underground, but much harder to conceive what might be 
meant by programming a robot to get to know a city.

The metaphor of "knowing the city" is helpful when understanding both the way in which 
an EM model is realised on a computer, and the philosophical stance it reflects on 
knowledge. As I look out of my study window at this moment, my view of the city of 
Coventry is limited to the roof of my neighbour's house and the sky above. In my 
imagination, I can trace the path from the house to the city, though the act of tracing this 
path is no part of my present direct experience. My knowledge of Coventry invokes the 
conjunction in my memory of all the direct experiences I have had over many years of 
different aspects of the city. There is no sense in which all these prior experiences can be 
taken as one experience, any more than I can be in more than one place at once. From this 
perspective, the 'perfect knowledge' on which we might aspire to base a classical 
computer-based model of the city is a mere chimera. What there is to be known of 
Coventry is more than I can ever experience, and my personal knowledge is established, 
maintained and revised dynamically through my ongoing interactions with it. Even my 
current limited view of the city is potentially open to such elaboration of knowledge, now 
that I notice the pigeon droppings on the crest of the roof, the silver sliver of a distant 
passing plane, and savour the taste of lukewarm lime juice cordial. As a source of direct 
experience, an EM model cannot compare with my everyday Coventry environment in its 
richness, but it offers access to potential experience and knowledge with which I can 
engage in similar ways. In effect, it supplies an interactive artefact with which to trace a 
history of experiences that is similar in character to our everyday experience of 
purposefully and accidentally developing our knowledge of a city. In this activity, the 
computer serves two significant roles: it serves as a physical artefact, similar to the 
artefacts of bricolage, that generate direct experiences to which the modeller can make a 
creative response, and at the same time it allows the modeller to record and elaborate 
experiences as they are encountered in 'the stream of thought'. In these respects, the 
computer offers support for managing knowledge as characterised as 'interaction with 
memory'.

1.2. An illustrative case-study

Empirical Modelling principles and tools were not initially developed with the general 
characteristics of EM models, as described above, in mind. Our present understanding of 
the characteristics of EM models was arrived at after first appreciating the difficulties of 
providing a formal semantics for such models within the classical theoretical framework 
for computation. Two sources of inspiration have been particularly significant in reaching 
this understanding: a large body of practical work on model-building and tool 
development that has been carried out largely by computer science students at Warwick 
over the last 15 years, and the philosophical writings of William James [13]. This 
subsection will give a brief sketch of practical EM with reference to a model of a lift 
initially developed in a summer vacation project in 1994. (The lift model and the EDEN 
interpreter which is needed to run the model can be downloaded from the EM website Ð 
see [17] and [18]). The next section will then discuss EM in relation to William James's 
philosophic attitude of Radical Empiricism [13].

Figure 1 depicts our EM model of the lift as it might be displayed in one particular state. 
The lift car appears on the left, and the five floors that the lift visits on the right. The stick 
figures are used to indicate the current locations of lift users: a label attached to the head 
of each stick figure identifies the list of users at each location. The bold vertical lines 
between the lift shaft and the floors represent the status of the doors between the lift car 
and each floor: the door on the fourth floor is currently open. The small boxes to the right 
of the floors and to the left of the lift car are call buttons, and those that are currently 
selected are indicated in bold.

To interpret Figure 1 as associated with a direct experience of an actual lift, it is helpful 
to think of looking at a glass lift at the side of a five storey building. Whether this is 
representative of the direct experience that is of primary interest to the modeller depends 
upon the role in which the modeller interacts. Examples of possible roles that the 
modeller might play, possibly concurrently, include: a lift user, designer or analyst, or 
perhaps even a story teller for whom a lift is a significant location (e.g. as the venue for 
murder in a detective story). In practice, if we have any familiarity with lifts (even if we 
are in the position of user X in Figure 1, with limited access to the actual current status of 
the lift), we bring to our interaction with them a general concept that there are floors and 
users and a lift car that moves vertically between them, and may be expected to visit the 
ground floor at some stage in response to pressing the call button.

 

In our modelling tool (EDEN), the current status of the lift is determined by the current 
values of its characteristic observables. The names of some of the key observables, as 
they are recorded in our model-building interpreter, are:

_liftfloor	- where is the lift? [on which floor: 1-5]
_open3	- is the door open at level 3?
_car2		- is button 2 in the car selected?
_button4	- is the call button on floor 4 selected?
locX		- where is user X?  [0: in lift, 1-5: on which floor, 7: nowhere]
_inliftX	- is user X in the lift?	
_destX	- where is user X intending to go?

The values of most of these observables are directly reflected in features of the display. 
For instance, _liftfloor determines the position of the lift, _open3 determines how 
the line that represent the door on floor 3 is depicted and _button4 determines whether 
the call button on floor 4 is highlighted. It is characteristic of modelling with EDEN that 
these features of the display can themselves be construed as observables whose values are 
specified by definitions resembling the definitions that relate the cells of a spreadsheet. 
Taken together, the definitions of observables, whether explicitly by values, or implicitly 
via formulae, form a 'definitive script' that specifies the current state of the model. The 
term 'definitive notation' is used to refer to the underlying notations used to formulate a 
script. EDEN includes a definitive notation for defining the screen layout and for defining 
planar line drawings. The use of these notations is illustrated in Figure 1. 
	
The definitions in a script express expectations about how the values of observables are 
linked where 'atomic' changes are concerned. For instance, when the lift moves, anyone 
in the lift moves with it as part of one and the same state change. The relations linking 
these changes are called 'dependencies' between observables. For example, using carpos 
and posX respectively to represent the positions of the lift car and of user X (as depicted 
by 2D coordinates in Figure 1), definitions of the following general type express 
dependencies:

	carpos is {liftshaftL , (_liftfloor-1) * floorheight}
	posX is if _inliftX then carpos else floorX
	_inliftX is (locX == 0)
	
The use of a definitive script to represent state-as-experienced is the fundamental 
technique by which a modeller constructs environments for interaction in EM. Such an 
environment is typically oriented towards expressing the perspective of a particular agent 
where both observation and interaction are concerned. In this context, an agent is 
anything that might be deemed to be responsible for changing the current state of the EM 
model. The archetypal agent is the modeller who acts to change the state of the model 
through manually entering redefinitions of observables into the EDEN input window. As 
will be discussed and illustrated later, the lift users and the lift itself can also be viewed 
and in various different ways animated as agents with some autonomous capacity to 
change state. In viewing the EM model, it is important to understand its primary role as a 
representation of state-as-experienced, rather than merely as a vehicle for automating 
conventional lift behaviour. To return to the analogy introduced above, animated 
behaviours in EM are in the first instance like bus tours of the city: by default, on such a 
tour, we observe the city according to a pre-programmed plan, but there is nothing in 
principle to prevent us from leaving the organised party and continuing our exploration 
independently. 

The idea that the lift model captures state-as-experienced can only be appreciated by 
considering the possible modes of interaction with it that are open to the modeller. To 
understand this, it is helpful to reflect on the distinction between the abstract interaction 
that is involved in using the underground and that involved in making the same journey 
above ground. In the former context, the traveller can get by by pattern matching, relying 
on the way in which the underground environment itself is built to encode knowledge 
about locations and directions in place names, keywords and icons. In the latter context, 
the personal knowledge of the traveller Ð their ability to recognise landmarks, to maintain 
a sense of direction, to devise strategies for crossing roads, to be able to consult other 
people Ð is paramount: though the environment itself has embedded knowledge in the 
form of signposts and street-names, the logistics of their use are by no means fully 
systematised. In practice, both opportunistic and exploratory human interpretation and 
pre-coded knowledge play their part in navigation in a city environment. In a similar way, 
interaction with the EM model of the lift can either give prominence to how human 
intelligence informs state transition or support routine and automatic interaction. What is 
more, commitment to changing the model is not necessarily required in playing these 
different interactive roles, and either can be highlighted according to purpose.

The openness of the modeller's agency in EM is the platform for intelligent interaction. 
The modeller is free to redefine any of the observables in the script subject only to being 
able to interpret the consequences of such redefinition. As a simple example, a 
redefinition such as

carpos is {liftshaftL , 0}

relocates the lift car at the base of the lift shaft as if it were out of service. Note that this 
redefinition in some sense 'deals intelligently with user Z': in the context of Figure 1, it 
would move user Z within the lift no matter where the lift was placed. Actions of this 
open-ended nature are analogous to the unconstrained actions that the traveller can make 
in the city environment, that may result in reaching the destination inadvertently, might 
lead to getting lost or to being run over.

To illustrate how knowledge about meaningful interaction can be exposed without being 
encoded in an EM model, the lift model has been adapted and used to animate the 
unusual instance of real-life lift use described in Box 1. For this purpose, the modeller 
has only to compile a sequence of elementary redefinitions of observables to reflect the 
actions of the users and the lift as they occur. By playing through this sequence of 
redefinitions, the modeller can visit each situation as it arises, and reflect upon the 
perceptions and motivations of the various agents involved. The modeller can then 
construct a narrative to document their informal understanding of the scenario. The EDEN 
interpreter provides a procedural construct for playing out this sequence of redefinitions 
as if they were being entered one-by-one by the modeller. Such a sequence can be 
accompanied by a commentary that gives more insight into the complex combination of 
observation and logic that guides each lift user's actions.

By way of context, the events described in this scenario occurred some fifteen years ago, 
and have been a source of puzzlement to the author ever since. Constructing the model 
has finally enabled me to reconcile my assured personal knowledge of the circumstances 
of the interaction with a plausible objective account of the behaviour of three sane lift 
users. To dramatise the distinction between these two perspectives on the events, I shall 
complement the personal account in Box 1 with a further analysis in which my role is 
described in the third person, as that of 'user I'. 



I am at a conference in the Netherlands.
I arrive late at night and hardly notice where my room is.
Next morning, I notice that my room is on the top floor.
I walk down to breakfast thinking about my talk later on.

After breakfast I meet two other delegates X and Y.
We get in the lift to return to our rooms.
X presses the button for floor 4.
Y says he is on the floor above X, and selects floor 5.
Since the top button is selected, I donÕt press a button.

We talk as we ascend. The lift stops. The door opens.
The floor numbers arenÕt clearly marked.
I say to X Ôthis must be floor 4' Ð he gets out.
Y and I carry on talking.

When the lift next stops, the floor is still unclear.
I say to Y ÔX is on the floor below you; this is your floorÕ.
Y gets out. I think something is not quite right.
I think Ôis this the top floor?Õ and Ôshould I get out?Õ.
IÕm unsure, but notice that the button for floor 5 is still lit.
I proceed to the top floor which is the next floor, floor 5.
When I get out of the lift, I canÕt find my room.

ThereÕs no room where my room is on floor 5.
I walk down to floor 4, and pass Y on his way to floor 5.
When I reach floor 4, I meet X coming up from floor 3

How did I manage to get all three of us to the wrong floor?


Box 1 : Travails in a lift

Figure 2 depicts a critical moment in the scenario, at which it becomes clear to user I that 
there is some conflict between what I observes and the way in which I had conceived X 
and Y's interaction with the lift up to this point. I supposes that his room is on the top 
floor, and realises that the lift has not yet reached the top floor. On the other hand, on 
entering the lift, I observed the selection of just two buttons by co-users X and Y, and 
noted that one of these corresponded to the top floor. It would seem from this that I 
should have got out at the same time as Y. It was just after Y got out of the lift that I 
realised that perhaps he had misled X, and that the lift had in fact stopped at floor 3 when 
called by an impatient user Z. At that point, it was apparent to I that both X and Y had got 
out at the wrong floors, but not that he himself was also about to travel upwards to the 
wrong floor. The extra detail of the profile of the building that has been introduced in 
Figure 2 represents the other element in the situation by which I himself was misled into 
supposing that his room was on floor 5 rather than floor 4: in the vicinity of I's room, 
floor 4 was the top floor. The pragmatic nature of the decision that I makes at this point, 
so different in character from the pure reasoning that is sometimes imputed to human 
agents (cf. the Mensa problem discussed in [1]), is emphasised by the real time 
constraints on decision imposed by the lift itself. 


Figure 2 can be interpreted as the particular specialisation of the vanilla model of the lift 
depicted in Figure 1 that I ideally should have had in mind in using the lift. In practice, I 
was not able to construct this model in real-time, and acted with a different conception of 
the situation. A crucial aspect of the EM model is that it relies essentially upon physical 
artefacts to account faithfully for I's misunderstanding. The geometric content of the 
model, limited as it is as far as realistic details of the physical environment are concerned, 
has rich experiential significance. It is to remembered observation of physical models 
such as are depicted in Figures 1 and 2 that we appeal in using a lift, and the experience 
of interacting with these models that exposes their role in representing our knowledge of 
lifts. Needless to say, many attempts have been made to abstract the essence from such 
experiential representations in logicist approaches to knowledge representation (cf. [10]). 
In practice, there are many other pertinent issues connected with the experience of lift use 
that influence the behaviour of users, many of which are only tacitly acknowledged even 
in our account of the scenario described in Box 1. These issues include: the nature of the 
experience Ð the time it takes a lift to move between one floor and the next, the relative 
time it takes to ascend 3 rather than 4 floors, the acceleration of the lift car; the nature of 
the auxiliary observation that might have affected user responses, as determined by the 
status of the buttons of the lift, the visual cues to distinguish different floors, whether and 
how the users' room keys and the lift car buttons were numbered; the implicit 
conventions framing the lift design and use, such as whether the ground floor was 
deemed to be floor 0 or floor 1, the fact that the buttons in the lift were ordered top to 
bottom in a 1-dimensional array in correspondence with the five floors, or that a call 
button in the lift was visibly de-selected when the door opened at the corresponding floor.

As the above discussion illustrates, the EM model of the lift offers insight into real world 
state-transitions about which the automation of the perfect lift user can give no 
information. Interaction with Figure 2 offers a plausible explanation for corporate 
behaviour that is absurdly different from the ideal outcome to be expected, but it also 
shows that Ð taking appropriate observational and experiential factors into account Ð the 
individual actions of the users are not so very far from sensible as they might appear. 
Logic alone cannot account for the scenarios that lie near to normality 'in the 
neighbourhood of sense'. In exploring such scenarios fully, there is no alternative but to 
allow the modeller to engage as freely as possible in the state-changing activity. In the 
spirit of Gooding's definition of the term, an EM model can be interpreted as a construal 
[11]. In constructing such a model using EM principles, our aim is to embody the patterns 
of observables, dependency and agency that we deem to be characteristic of a situation. 
The product of this activity typically has a personal, provisional and particular character 
that is intimately connected with the ways in which we choose to interact with it. In all 
these respects, it is unlike a formal specification.

In developing a construal, the modeller is not only concerned with modelling state change 
as it is observed from an external perspective. A further analysis of the sequence of 
redefinitions that are made in realising the scenario in Box 1 shows how certain groups of 
redefinitions can be attributed to different agents. The movement of the lift and the 
manipulation of the lift doors are part of the automatic behaviour of the lift system itself. 
Entering and leaving the lift and the selection of call buttons inside and outside the lift are 
actions for which the lift users are responsible. To refine the model so that it better 
reflects our understanding of normal lift operation, it is appropriate to distinguish 
between realistic transitions and transitions that are pure fantasy, such as might involve 
simulating a user entering the lift before the door is open, or the lift jumping between 
floors. EM includes techniques for enriching the lift model so as to give special 
prominence to construals suited to different perspectives and purposes, some of which 
may include automated behaviour, or reflect several viewpoints concurrently. It is 
important to realise that the specialisation of an EM model does not involve excluding 
singular actions on the part of the modeller. In some circumstances, such as when the lift 
cable breaks, it may be appropriate for the lift to exhibit abnormal behaviours that can 
only be simulated by the intervention of the modeller in the role of a super-agent. The 
EDEN interpreter is designed to support the opportunistic interaction by the modeller that 
is required for this purpose, but whether such privileges for interaction are exercised is 
entirely at the modeller's discretion. A brief review of some relevant EM techniques is 
given here Ð for more details, consult the references at [17].

In elaborating construals, a significant role is played by an informal special-purpose 
notation, called LSD, that can be used to classify the observables in a situation with 
respect to an agent. The aim of such a classification is to document the roles that 
observables play in determining the interaction of an agent. Listing 1 is an LSD account 
of a person X in the role of a prototypical user of the lift depicted in Figure 2:

agent person (X) {
	state	 loc[X]
	role liftuser {
		state 	dest[X]
		oracle 	open[*], dest[X], loc[X], pos[X]
		handle 	loc[X], dest[X]
		derivate 
			pos[X] is if loc[X]==0 then liftfloor else loc[X], 
			LIVEliftuser[X] is 0<=loc[X]<=5
		protocol
			loc[X]==0 and open[dest[X]] and pos[X] == dest[X]
					-> loc[X] = pos[X],
			true	-> dest[X] = i ,				(1<=i<=5)
			...
	}
	...
}

Listing 1: An LSD account of a prototypical lift user

In this account, the observables loc[X] and dest[X] refer to the location and 
destination floor of user X respectively. The values of these observables are defined 
according to the same conventions used to define locX and destX in the scripts 
discussed above. The location of X is classified as a state for X as a person, since 
observation of this location is meaningless in the absence of X, and the destination floor 
of X is classified as a state for X in the role of lift user, since the concept of X's 
destination floor is meaningful only when X acts in this role. Whether or not X is 
currently playing the liftuser role is reflected in the boolean value of the special 
observable LIVEliftuser[X], which is true or false according to the current value of 
loc[X]. As a lift user, X has access to certain observables as oracles, either in the sense 
that she can observe them directly, or that she can observe them at certain times, and may 
retain some notion of their current values. She typically knows her destination floor. She 
also knows whether she is in the lift, or at a floor, as determined by the oracle loc[X]. If 
X is in the lift, she has some notion of what floor the lift is currently at, as represented by 
the observable pos[X]. The observable pos[X] is classified as a derivate because of the 
dependency that defines it in terms of the current lift position and the value of loc[X]. 
The observables over which X can conditionally exercise control are classified as 
handles: they include X's location and destination. The conditions under which X can 
redefine these observables are set out in her protocol. The two example privileges 
specified for X indicate that if X is in the lift, the lift is at her destination floor and the 
door is open, then X can step out of the lift, and that moreover X can change her mind 
about her destination floor at any time.

An LSD account is not in general to be interpreted as a specification, but rather as 
documenting the characteristics of observables as they are experienced by the modeller in 
the real situation, and (if the EM model is sufficiently convincing) as they are 
experienced in the associated EM model. In effect, an LSD account is intended to 
complement an artefact with which the modeller can interact, and is not to be viewed as 
an alternative form of representation. An LSD account can play a significant role in 
guiding the development of several different kinds of EM model for a multi-agent system 
such as a lift. These include:

Concurrent modelling of user perspectives on lift use. An LSD account can be seen as 
describing the system as viewed from the perspectives of its various users. These views 
can then supply the basis for a distributed EM model in which the personal construal of 
each agent is modelled as a client in a client-server configuration, and the corporate 
behaviour of the agents is developed by managing their interaction via the server. For 
instance, using a distributed variant of the EDEN interpreter, it would be possible to set 
up an EM model similar to that depicted in Figure 2 on a server, and to create EM models 
on four clients to recreate the interaction of the lift users X, Y, Z and I in the scenario as 
it appears from their four different perspectives. In such a model, there will be nothing 
that corresponds to the lift users X, Y and I from the perspective of user Z, since Z 
merely presses the call button on floor 3 then descends to floor 2. The models for X, Y 
and I will be based on different variants of the prototypical lift user account given above. 
For instance, the oracles dest[X], dest[Y] and dest[I] have to be construed quite 
differently: user X has a correct perception that dest[X] is 4, but (when encouraged by 
I), supposes that pos[X] is 4 when it is in fact 3; user Y has an oracle to dest[X] which 
takes its value from the floor on which X gets out, and interprets dest[Y] as a derivate 
defined by dest[X]+1; user I has the number of the top floor of the building as an 
additional observable topfloor and interprets the oracle dest[I] as a derivate defined by 
topfloor. A distributed model of this nature can provide a more vivid reconstruction of 
a scenario than the basic EM model depicted in Figure 2, as has been illustrated 
elsewhere in our previous research in the reconstruction of historic railway accident 
developed by Sun [4]. 

Modelling the lift as a reactive system. An LSD account can be used in analysing the 
stimulus-response behaviour of automatic agents as they are construed to respond and act 
upon observables in their environment. Activity of this nature has a fundamental role in 
experimental science. It is also an important aspect of reactive systems development, 
where it is associated with the exploration and specification of context that precedes the 
design of control software. In applying EM principles to such activities, the most 
significant feature is making the connection between model-building from a personal 
subjective perspective and what is interpreted as objective observation of a system by an 
external observer. A full discussion of the technical issues involved is beyond the scope 
of this paper Ð it has been one of the primary concerns of the EM project as a whole (cf. 
[1,4,6,15]). The basic concept is that of treating the activity of automatic agents as it were 
being carried out by a human agent with the appropriate perceptions and state-changing 
capabilities. By way of illustration, the lift system can be viewed as automatically 
carrying out the sequences of redefinitions required to reset the call buttons and open and 
close the doors on visiting a floor to which it is called. The stimulus for this operation is 
the presence of the lift car at a floor that is associated with a selected call button. An LSD 
account of this role of the lift system would (for instance) identify oracles Ð such as the 
status of call buttons, handles Ð such as the status of the doors, and include an action to 
manipulate the doors and call buttons appropriately in its protocol. A similar analysis can 
be used to develop a protocol to prescribe the motion of the lift. The EDEN interpreter 
includes features, such as procedures that are triggered by changes to specified 
observables, that can be introduced into the EM model of the lift to automate the lift 
system protocols. Such features make it possible to implement a lift simulation through 
an incremental process of extending the EM model in which the modeller's role involves 
shaping the behaviour to accord with realistic observation and interaction. The delay after 
which the doors close can be adjusted, for instance, and the selection of buttons by lift 
users simulated by direct mouse actions. The significant point here is that the LSD 
account has no formal operational semantics, but documents actual interaction with a 
computer-based artefact for which behaviours can be developed in much the same 
incremental and empirical fashion that an engineer might construct a prototype. In this 
process of empirical refinement of the EM model, the scope for extension is open: simple 
extensions that feature more realistic lift motion, implement a lift scheduling algorithm, 
and introduce prototypical users based on the LSD account in Listing 1 can be found in 
the liftBeynon2003 directory of the EM archive [18]. Further extensions for this 
model might involve introducing greater physical realism by way of modelling lift car 
velocity and acceleration, adding 3-dimensional visualisation, or linking the model to 
special-purpose hardware that could simulate the impact on the user of forces generated 
by the lift motion.

The above description and illustration of EM sets the scene for the discussion of Radical 
Empiricism that follows. To fully appreciate this discussion, it is most helpful to have 
some experience of the nature of EM as a practical activity. Without such experience, it is 
difficult to appreciate the conceptual distinction between traditional computer programs 
and EM models that is described in section 1.1. It is in particular important to realise that 
all the lift models discussed in this section are to be regarded as part of a single open-
ended conceptual process of exploration that can be seen as resembling the exploration of 
a city. Each model is informally associated with a particular way of viewing a lift 
situation and of organising the transitions between one situation and the next, but each is 
apprehended Ð even when, left to itself, it is executing a particular pattern of behaviour Ð 
state-by-state, in such a way that the modeller can choose to intervene to redirect the 
experience perhaps with a view to its reinterpretation. In this respect, the individuality of 
the different lift models is not defined objectively, but only with reference to what 
experiences are coherent for the individual modeller.

2. Radical Empiricism

William James's Essays in Radical Empiricism [13] was first published in 1912. The 
potential relevance of James's 'philosophic attitude' to a discussion of alternatives to a 
logicist framework for knowledge representation is apparent throughout these essays. In 
'The World of Pure Experience', for instance, when discussing the philosophic 
atmosphere of his time, James refers to 'a feeling that [the extant school-solutions] are too 
abstract and academic', and goes on to write: 'Life is confused and superabundant, and 
what the younger generation appears to crave is more of the temperament of life in its 
philosophy, even though it were at some cost of logical rigor and of formal purity'. 
Empirical Modelling is motivated by a perceived need to develop methods of computer-
based modelling that can do more justice to life in its confusion and superabundance than 
can the rational formal accounts of agent interaction that underlie typical computer 
programs. The distinction between these different views of human agency has been 
illustrated above when contrasting the farcical scenario of lift use associated with Box 1 
with the prosaic and predictable behaviour that is attributed to prototypical lift users in 
Listing 1. This section explores other respects in which the principles and techniques of 
EM can be related to thinking developed by James in [13]. Our overall aim is to make it 
plausible that James's philosophical stance supplies a foundational framework in which to 
examine issues that seem paradoxical from more conventional philosophical perspectives.

2.1. Philosophical foundations for EM

Some background motivation for our discussion can be found by thinking in general 
terms about what kind of philosophical foundations are appropriate for EM. One of the 
most characteristic activities in EM is the construction of a computer-based artefact (for 
instance, the lift model in Figure 2) that embodies patterns of observables, dependencies 
and agency that can be identified in an external situation to which the artefact refers (for 
instance, the specific use of the actual lift described in the scenario in Box 1). From a 
traditional computer science viewpoint, where there is an underlying presumption that all 
computer-based modelling can be accounted for by using the universal abstractions that 
rest ultimately upon the classical theory of computation, it is usual to propose that EM 
reduces to classical programming through a correspondence of the following general 
kind: an observable is a procedural variable, a dependency is a constraint relation, an 
agent is a sophisticated abstraction such as an active object. The renunciation of each of 
these proposed reductions has been the focus of special attention in our previous work 
(see for instance: the discussion of variables in [5], of constraints in [15] and of agency in 
[1]). Broadly, our counter to this suggestion is similar in all three cases: that the notion of 
'observable' refers to a feature of a situation that is experienced as having an identity and 
current status or value; that a 'dependency' is more than a perceived abstract relationship 
between values of variables and expresses the modeller's expectations about the 
immediate consequences of changing the values of observables in a situation; that 
'agency' entails a potential for action that is of a truly experimental character, in that the 
possibility of taking the action has not been preconceived, and that no prior commitment 
to the possible interpretation of its consequences has been made. What links each of these 
counterproposals is the context in which our interaction with the computer model is 
conceived to be occurring: a context resembling that through which I as-of-now 
experience the city of Coventry through the sight of my neighbour's house and the sound 
of his lawnmower.

Our rejection of a conventional interpretation for EM has another significant element that 
is closely linked to Brian Cantwell Smith's critique of classical computational semantics 
in [9]. The proposed reduction of an EM model to a classical program purports to 
attribute an abstract behaviour to the EM model, in accordance with the ways in which 
variables, constraints and active objects might be used to specify behaviours in a 
conventional approach to model-building. The development of an EM model does not 
rely upon the identification of an abstract behaviour that can be embedded into the 
environment through formal symbolic associations. This is to revisit our previous 
observation, that an EM model does not in the first instance specify an activity, like 
travelling about a city using the underground, for which Ð thanks to the traveller's training 
in symbol recognition, and the careful engineering and signposting of the underground 
environment Ð limited experience of the city is required. On the contrary, the EM model 
offers itself to the modeller as a state to be experienced, where the correspondence 
between the features of the model and those of its referent are to be directly established, 
explored and enhanced through interaction. This perspective has to be understood with 
reference to a philosophical position that assumes no given absolute knowledge, in the 
same spirit that (in my current context) I cannot be absolutely sure that all I remember of 
the city of Coventry will be there to experience when I set off to visit the railway station.

As Cantwell Smith observes in [9], the inadequacy of the classical view of computation is 
conspicuously exposed in emerging computing practice. The wide range of applications 
for EM principles that we have identified to date highlights both the potential of EM and 
the challenge of understanding its semantics. The pragmatic view of the semantics of 
modelling 'real-world objects' that serves well enough in traditional engineering design 
can to some extent be adapted to an exercise such as using EM to model an actual lift. 
The concept of a direct correspondence between experience of an EM model and the 
experience of the real-world situation it represents requires more justification when the 
EM model is a spreadsheet, and the situation a financial scenario. It is also more difficult 
to argue for a direct correspondence between observables for an EM model to represent a 
lift that is under design and has yet to be built. Further complications arise if we consider 
the status of the EM model of a lift that we might create in a virtual reality, where the 
relationships between observables are no longer subject to familiar physical laws and 
constraints. EM has been used to build models that have the experiential characteristics of 
the data structures (such as the heap) that lie behind standard algorithms (such as 
heapsort), and thereby to generate an environment in which to explore the design of 
algorithms [6]. In other contexts, it is apparent that the use of EM principles is not 
directly concerned with issues of external representation. For instance, our tools now 
enable us to design new definitive notations Ð in particular, for graphics Ð within the 
same paradigm that we use to construct models, and Ð for such a notation Ð successful 
design is concerned with how the syntax of the new notation affects the correspondence 
between the structure of a script and the graphical image that it produces on the computer 
screen.

A feature common to all these applications of EM is the construction of a computer-based 
artefact that in some respects has the qualities of an instrument (cf. [3]). The term 
'instrument' is used here to express the idea that there is a reliable correlation between the 
interactions of the modeller with artefact and its associated changes of state. This 
correlation may be explicitly engineered by the modeller, or learned through skill 
acquired in mastery of the instrument. Explicit engineering is prominent in the case of a 
financial instrument such as spreadsheet, where the definitions of cells are contrived to 
express known relationships between observables. Skilful interaction is more prominent 
in relation to a musical instrument, such as the violin, where the tiniest nuances in the 
movement of the bow can be used to control the sound generated. In broad terms, the 
account of EM that we have sought to develop in our project represents EM models as 
instruments simultaneously under development and in use by the modeller, in which both 
explicit engineering and skilful interaction have a role to play. Through explicit 
formulation of dependencies and through experimental redefinition, the modeller 
develops an understanding of how interactions with an EM model reliably effect changes 
to its state. In this account of EM, the interaction with the instrument does not acquire 
meaning as a result of a complex abstract process of off-line decoding; it directly evokes 
a parallel experience because of the perceived similarity between the effects of 
interaction with the instrument and the familiar effects of interaction in another context. It 
is in just this manner that the movement of geometric elements of the simple drawing in 
Figure 2 evokes familiar interactions with an actual lift.

The correspondence between one experience and another that underpins EM is different 
in character from the realistic modelling of behaviour and appearance that is 
commonplace in routine prototyping of software and engineering systems. For instance, 
our primary concern is not with simulating the lift dynamics as accurately as possible by 
analysing the forces acting on the lift in detail and applying Newton's laws, nor with 
developing a virtual reality model that imitates the user's visual and sensory experience as 
faithfully as possible. As the lift model illustrates, EM can serve to establish such 
similarities, but the connection between an EM model and the situation to which it refers 
is more direct and primitive in nature. The correspondence between interactions with the 
EM model and interactions with its referent is itself a matter of immediate experience, 
subject to confirmation, exploration or possibly even refutation by the modeller 'as of 
now'. A correspondence of this nature is rooted in the notions of 'observable', 
'dependency' and 'agency' as they have been discussed above: it has a concrete and 
dynamic rather than abstract and static quality. In immediate experience, the agency of 
the modeller has the capacity to confirm or confound expectations, to create or destroy 
observables, to make or break dependencies. In these respects, it resembles the agency of 
the experimental scientist, whose actions can be used to test her construal.

The potential role that language and symbolic conventions can play in this context is a 
delicate issue. It is self-evident that our apprehension of correspondences in immediate 
experience can be mediated by language, as when we respond to the value in the column 
of the spreadsheet that is headed 'balance' or 'profit'. It is quite another matter to argue Ð 
as some philosophical traditions appear to do Ð that language plays an essential role in all 
correspondences between experiences. A core idea in EM is that not all correspondences 
between one experience and another can be established by symbolic conventions Ð at 
some level, correspondences must be made through direct experience without reference 
to language. The expectations of the experimental scientist may well entail much 
sophisticated theory, but the primitive correspondences that provide the grounding for 
such theory are arguably beyond words and equations. Experience in applying EM 
principles and tools suggest the further hypothesis that agency, dependency and 
observation have a fundamental role in such primitive correspondences.

2.2. Radical Empiricism from an EM perspective

Experience of EM, and reflection about its semantics, endorses a philosophical position 
that has strong points of connection with William James's Radical Empiricism. The 
primitive correspondences between experiences at the core of EM are necessarily 
personal experiences, and in the first instance are associated with subjective and 
provisional knowledge. This establishes priorities that are in line with those of RE: to 
account for logic and theory in terms of observation and experiment, rather than to 
account for observation and experiment in terms of logic and theory. Practical experience 
of EM reinforces this perspective, demonstrating how EM can be accompanied by 
transitions from personal to public, subjective to objective, and provisional to assured (cf. 
[1] for more discussion of these issues). This section discusses how some key ideas of RE 
are helpful in elaborating on what is involved in EM.

In [13], James remarks that, in RE, "the relations that connect experiences must 
themselves be experienced relations, and any kind of relation experienced must be 
accounted as 'real' as anything else in the system". One of his primary concerns is that 
traditional empiricists emphasise the disjunction in experience "leaving things 
permanently disjoined", whilst the rationalists remedy this disjunction "by their 
Absolutes ..., or whatever other fictitious agencies of union they may have employed". 
For James, both conjunctive and disjunctive relations should be deemed to be given in 
experience. Amongst these, he includes "the most intimate of conjunctive relations, the 
passing of one experience into another when they belong to the same self".

James's outlook helps to explain the difficulty we have encountered in understanding and 
communicating the nature of EM models. The typical computer scientist has a natural 
desire to attribute a characteristic set of discrete states and behaviours to the EM model of 
the lift in Figure 1, and to view it as a structure or system. In the mind of the modeller, 
the character of the EM model is more elusive. We can experience the state of the lift 
model as it is depicted in Figure 1, but interpret this in relation to all the other 
experiences of the states that have brought the model to this point. By the same token, 
what we experience in the current state of the model stands in an open, yet to be 
explored, relation to other experiences we can get by interacting with the model. What we 
understand by the model is an unfolding conjunction of experiences that cannot be 
circumscribed but is apprehended as a single relation.

An important aspect of James's analysis is the implicit emphasis it places on the 
authenticity of the personal experience of the observer. This distances RE from the 
perspective with which empiricism is ordinarily associated, where the 'reality of the given 
world' is seen as the primary source of knowledge. As our practical experience of EM has 
shown, it is implausible that we can give a good account of EM without regarding reality 
as constructed by experience. Though it is convenient in describing EM to adopt 
everyday terminology, and speak of 'the modeller's state of mind', and 'the real-world 
situation', this should not be understood as subscribing to a Descartian dualism. This 
accords with James's observation that "subjectivity and objectivity are affairs not of what 
an experience is aboriginally made of, but of its classification. Classifications depend on 
our temporary purposes ... ". In observing the development of the EM model of the lift, 
we are led to think quite as much about the evolving state of mind of the modeller as 
about realistic changes to the state of the actual lift, whether these take the form of a 
movement from one floor to the next, or adding labels to the call buttons in the lift car. 
This is keeping with our perception in EM that, contrary to the received view that a 
computer model of lift should only Ð perhaps even can only Ð be constructed with a 
specific goal and behaviour in mind, our EM model of the lift is somewhat neutral to 
purpose. Certainly, the experiences of the lift model that would be generated through the 
interaction of modellers acting in the roles of lift users, lift designers, lift analysts, or 
story tellers would be quite different, and reflect many different kinds of conjunctive 
relation in their minds.

Of particular relevance to EM is James's contention that "the first great pitfall from which 
[RE] will save us is an artificial conception of the relations between knower and known" 
[13]. This issue relates directly to the discussion above concerning how correspondences 
between experiences are established. The artificiality to which James alludes here stems 
from what is perceived by other philosophers as the 'indefeasibly dualistic' structure of 
experience [13]. For James, there is no such duality: "All the while, in the very bosom of 
the finite experience, every conjunction required to make the relation intelligible is given 
in full" [13].

James's philosophical position is helpful in understanding the nature of EM. As has been 
argued above, the fact that the EM model of a lift stands in a special relation to an actual 
lift (viz. in what James characterises as 'the relation between knower and known' [13]) is 
not illuminated by classifying the one as a 'mental model' and the other as a 'real-world 
object'. What matters is that they are two portions of experience that are related in a way 
that is itself experienced by the modeller (cf. [13]). This leads us to a characterisation of 
the modeller that matches our experience of EM activity well: that of an agent who 
generates an experience that knows another. In the same spirit, EM principles and tools 
can be seen as assisting the generation of this experience. In this context, our choice of 
the word 'model' belies our faithfulness to James's account of knowing, as it is commonly 
associated with what James rejects by way of "representative theories".

The simplicity of the relation of knowing, as James describes it, is consistent with our 
experience of how readily EM models can be combined and reinterpreted. James rejects 
the duplicity of experience that distinguishes 'consciousness' from 'content' : "Experience 
... has no such inner duplicity; and the separation of it into consciousness and content 
comes, not by way of subtraction, but by way of addition Ð the addition, to a given 
concrete piece of it, of other sets of experiences, in connection with which severally its 
use or function may be of different kinds". This explanation accounts for the way in 
which the object of study in one exercise in EM seamlessly becomes a part of the EM 
model in another; for the fact that an EM artefact, like an artefact in bricolage, can be 
developed without a referent in mind; for the ambiguity about what features of an EM 
model serve a significant representational role (cf. the bold lines indicating the lift doors, 
the size of the call button panel in the lift, the presence of the roofs in Figure 2). In each 
of these contexts, the precise character of the EM model is only shaped by the nature of 
the interaction of the modeller as it unfolds at her discretion Ð in particular, by how the 
values of observables are interpreted and changed, and how these changes to observables 
are interpreted. By way of illustration, a student who used EM to simulate bread baking 
in an oven chose to represent the oven by borrowing a simple line drawing to represent 
the floor plan of a room (see roomviewerYung1991 at [18]), and labelled his 
simulation by substituting 'Bread baking simulation' for a warning message that notified 
the user when a table obstructed the door. This meant that the identifying label could be 
changed by relocating an invisible table Ð a functionality that the student did not 
document, and was unacknowledged in his personal interpretation of the model.

Though both RE and EM share a central concern for rooting knowledge in personal 
experience, their agendas are quite different, and explicit connections are hard to make. 
Despite this, RE is helpful in developing a deeper understanding of the primitive concepts 
of EM, for which formal logical foundations cannot be supplied.

James's account of pure experience provides a most appropriate setting in which to 
explain the notion of an observable in EM. The appropriate sense of being an observable 
in EM is 'having an identity' and 'having a value that can be directly experienced', where 
the term 'is directly experienced' refers to the capability of the human interpreter in the 
given context to apprehend immediately. There is a most significant distinction between 
this notion of an observable and what traditional empiricism might deem to be an 
observable. To the infant, the symbol '2' on the lift button is a mere pattern of sensation 
that signifies nothing; to a young child, it will be associated with the idea of 'some pair of 
objects'; to the competent lift user, it is known to refer to a specific floor. James's account 
of knowing dispenses with the idea that "seeing the symbol '2'", "thinking of two floors" 
and "imagining going to the second floor" are categorically different kinds of experience 
whose association in the mind of the lift user must be explained by "fictitious agencies of 
union". When situated in the lift, the lift user experiences the conjunctive relation that 
connects all three elements of his experience of the button labelled '2'. The uniform way 
in which observables are treated in an EM model, regardless of the level of sophistication 
of the observation involved so long only as it is immediate (cf. labelling the call button 
by 'II', or "The smallest prime number"), is consistent with James's outlook. The fact that 
in EM, as in life, the scope and significance of such observables is dynamically 
established as experience is acquired Ð perhaps from moment to moment, as a child might 
be taught in situ to connect the symbol '2' with a pair of floors Ð is also in keeping with 
James's conception of knowledge: "Why insist that knowing is a static relation out of 
time when it practically seems so much a function of our active life?" ([13], p.75).

In EM, there is an intimate relationship between the notions of dependency and agency. 
Dependency is in effect a way of binding together all the consequences that are deemed 
to be an integral part of a single action. In EM models, this concept is realised practically 
through modelling with definitive scripts (cf. [15]), where a typical atomic action 
involves redefining an observable or introducing / removing a cluster of observables. An 
analysis of the notion of dependency reveals a number of respects in which it is related to 
agency. Whether a dependency is identified in a particular context in general depends on 
the perspective and agency that the modeller has in mind. The position of the lift car 
might be seen as dependent on the position of the lift cable, but this view is entirely 
appropriate only if we assume an idealised mechanical model (e.g. making no allowance 
for the initial extension of the cable under load), discount the discrepancies between the 
exact positions of the lift car within the tolerances allowed in the design of the lift shaft, 
and decide not to model the behaviour of the lift cable at the atomic level. A naive 
mechanical model of the dependency may need to be modified if the possibility of 
thermal effects is admitted, and is no longer appropriate if the lift cable is presumed to 
become slack or to break. Nor is dependency necessarily associated with synchronisation 
of change; it has more to do with what we informally understand by causation, as when a 
doctor declares that a living person has been fatally wounded. This leads us to view 
dependency as framing what consequences of an action are an inevitable effect of the 
action, and cannot be allayed by the intervention of any other agent. There is also an 
important distinction between the notion of dependency that is often invoked in analysing 
concurrency in traditional procedural programs, which broadly relates to how the current 
value of a variable depends on previous assignments to other variables, and the 
indivisible characteristic of the dependency that in general binds many changes to one 
atomic change in EM.

There is no explicit reference to dependency in James's account of pure experience. 
However, amongst the conjunctive relations he lists "relations of activity, tying terms into 
series involving change, tendency, resistance and the causal order generally". One of the 
most significant features of EM is that it provides an alternative to the classical 
procedural model of state, where the values of variables are treated as discrete and 
independent: the dependencies in a script are not assertions about values alone, but 
embody expectations about responses to interaction that are themselves a part of state. To 
the extent that these expectations are relations between observables that are directly 
experienced, it seems appropriate to classify dependencies as conjunctive relations of 
activity. A plausible analogy likens the disjoint terms of traditional empiricism to the 
discrete variables of procedural programming, and the conjunctive relations of Radical 
Empiricism to the dependencies in modelling with definitive scripts. The elaborate 
programming mechanisms that are required to maintain dependencies between variables 
in a procedural representation of state echo James's comment about the complex ways in 
which rationalists remedy disjunction.

James's perspective on the relationships between transitions in experience is helpful in 
clarifying the interpretations of interactions and dependencies in EM models. In the EM 
model depicted in Figure 1, for instance, the lift is conceived as 'moving from one floor 
to the next' though there is no explicit attempt to model continuous lift motion. Within 
this naive lift model, the location of the lift user is thought of as defined by 'in the lift', 'on 
a particular floor' or 'not in the vicinity of the lift'. This 'user location' cannot be identified 
with an actual physical location, but is nonetheless deemed to be a meaningful 
observable. In this context, just exactly where the lift user is standing in Ð or in the 
vicinity of Ð the lift, or what posture they adopt, is not significant. From this, it may 
appear that the dependency by which "the position of the user is determined by that of the 
lift" is to be construed in a different way from the dependency that links the position of 
the lift car to that of the end of the supporting cable. The correlation between the 
movement of the end of the cable and that of the lift car is far more exact, and is easily 
interpreted with reference to atomic 'infinitesimal' change. To conceive the movement of 
the lift between one floor and the next (as modelled by _liftfloor++;) as an atomic 
change is to presume that no matter what the lift user does within the lift, they will be 
moved with the lift from one floor to the next. In the context, the relationship between the 
continuous motion of the lift and the discrete movement from floor to floor can be 
interpreted as illustrating what James identifies as substitution, whereby "an experience 
that knows another can figure as its representative, not in any quasi-miraculous 
'epistemological' sense, but in the definite practical sense of being its substitute in various 
operations, sometimes physical and sometimes mental, which lead us to its associates and 
results". The natural way in which these two experiences of a lift can co-exist subject to 
appropriate assumptions about the modeller's motivation is illustrated in the extension of 
the EM model mentioned above (cf. liftBeynon2003). For this purpose, we need only 
introduce a new observable liftfloorheight that (say) can assume integer values 
from 1 to 5*N (where N=10 for instance) corresponding to lift positions at or between 
floors, and define _liftfloor to be the integer part of liftfloorheight/N. Such a 
model is appropriate for many practical purposes, but would not serve for a murder 
mystery in which we might well be asked to imagine that a person ascending in the lift 
from floor 3 fails to arrive at floor 4, or meet the need the engineer may have to consider 
non-integral positions for _liftfloor for a lift that is malfunctioning.

As the above discussions indicate, the character of an EM model is moulded by the way 
that the modeller chooses to interact with it, and how this interaction reflects the 
modeller's evolving presumptions about the agency to be taken into consideration. In 
much the same way that each visit to Coventry serves both to fulfil familiar expectations 
and functions and to introduce what is changed or was previously unknown, interaction 
with an EM model involves both creation and use. In contrast, the formal specification of 
computer models requires a commitment to modes of agency and interpretation (i.e. in 
'creating' the model) that cannot be reappraised in the subsequent course of interaction 
with the model (i.e. in 'using' the model). This accounts for a fundamental difference in 
orientation between EM and traditional computer-based modelling. In the former context, 
our agenda is 'finding a good construal', for which the key question is 'given that this is 
what we're interested in achieving, what assumptions about agency in the world is it 
necessary and appropriate to make?'. In the latter context, our agenda is 'optimising our 
use of known resources for familiar purposes', for which the key question is 'given that 
this is the agency in the world, how do we best exploit it?'. The premise for the former 
agenda is ignorance of the world, and for the latter, knowledge of the world.

From a philosophical perspective, RE can be viewed as endorsing this shift in 
engineering priorities that EM promotes. In "The Experience of Activity", James [13] 
remarks: "... the healthy thing for philosophy is to leave off grubbing underground for 
what effects effectuation, or what makes actions act, and to try to solve the concrete 
questions of where effectuation in this world is located, of which things are the true 
causal agents there, and of what the more remote effects consist". The thrust of agent-
oriented analysis in EM is arguably well-aligned to James's recommended agenda in its 
concern for identifying agency, attributing state-change to agents and interpreting agent 
interaction in global state-based terms. In contrast to the mainstream traditions of 
research on agent-oriented modelling and programming, EM favours a concept of 
'agency' that is more than any circumscribed preconceived rationalised interaction, and is 
oriented towards a pragmatic dynamic shaping of construals. This emphasis is consistent 
with James's recommendation that, in examining 'the real facts of activity', and arbitrating 
between whether our actions are programmed by a higher authority, are an expression of 
free will, or emerge from the corporate behaviour of more primitive agents, we should 
evaluate our responses to the question "Whose is the real activity?" by asking "What will 
be the actual results?" [13]. In principle, EM provides a practical framework within 
which to tackle this agenda, supplying environments in which to explore 'possible 
construals' and to situate the negotiation of meaning.

As the discussion of the possible extensions of the EM lift model illustrates, the virtue of 
such an environment is that it is a source of experience that is rich to the point of 
incoherence. As James observes: "Experiences come on an enormous scale, and if we 
take them all together, they come in a chaos of incommensurable relations that we can 
not straighten out. We have to abstract different groups of them, and handle these 
separately if we are to talk of them at all." [13]. In the EM model of the lift, we can 
accommodate the possibility that the lift scenario described in Box 1 occurred not 
because Ð as I hypothesised Ð there was a user Z, but simply because the lift control was 
faulty. We can dramatise I's predicament in deciding whether to get out of the lift on floor 
4 without needing to resolve the logical inconsistencies in his perception and pursuing 
these to their contradictory conclusions. In contrast, conventional programming, like 
traditional empiricism, has no satisfactory way to handle the incompleteness of 
knowledge. Without such means, it cannot do justice to James's conception of ourselves 
as 'virtual knowers', or his observation that "To continue thinking unchallenged is, ninety 
nine times out of a hundred, our practical substitute for knowing in the completed sense."

Conclusion

The perspective on the nature of knowing that RE and EM endorse has direct practical 
relevance for current trends in knowledge management (cf. [7]). To date, the successful 
application of computers in management has relied to a large extent on exploiting what 
can be objectified and expressed in formal notations (as in a relational database, or an 
expert system). As we seek to make yet more sophisticated use of computer technology, 
and as this technology itself potentially embraces broader aspects of the total business 
experience, so the limitations of widely accepted philosophies of information science are 
being exposed. The problems we face are epitomised by the difficulties of negotiating 
ontologies and standardising formal representations and procedures for communication 
and re-use. This paper attributes these problems to the barriers that traditional 
philosophical frameworks for cognition and computation place between words and 
concepts and the experience that informs them.

The fundamental role that experience plays in informing concepts, as emphasised by both 
RE and EM, is patent in the everyday situations within which knowledge management 
has to function. Consider the experience that leads us as we grow up to identify one and 
the same entity as "a person", "a man", "a doctor", and "a paediatrician", or to appreciate 
the distinction between 'learning to speak French' and 'being French'. In attempting to 
capture such concepts in formal ontologies without reference to experience of artefacts, it 
is arguably impossible to do justice to the role of tacit knowledge, and to reflect the 
subtle nuances concerning the perceptibility, reproducibility and stability of the 
underpinning experience. As we aspire to provide computer support for 'experience 
management', and accommodate such broad perspectives on knowledge such as 
'mimetics' affords, there is ever more need to take explicit account of personal experience 
and to understand this in relation to our interaction with the natural world and with other 
people. RE in conjunction with EM potentially offers a philosophical and practical 
framework within which to give proper prominence both to direct experience and the 
personal stream of thought, and to the distinctions between knowledge as 'socially 
accepted' and knowledge as 'experientially validated'.

For the sceptical reader, a major intellectual objection to engaging with the thesis of this 
paper is that it represents RE and EM as essential alternative fundamental philosophical 
and computational perspectives. To acknowledge that concepts such as 'theory', 
'reasoning' and 'reality' have their significant place within this perspective is not enough 
to deflect such scepticism. In this context, a key issue is that we have become inured to 
interpreting our interactions with computing technology in a narrow sense that we would 
(arguably) not entertain as appropriate in relation to other experiences, such as playing or 
listening to a musical instrument. Indeed, the influence of computational theory on 
cognitive science has been such that there is a tendency towards construing all interaction 
as a form of computation. James's characterisation of the nature of knowing is of crucial 
significance in this connection: it identifies the relationship between one experience and 
another not as rationally apprehended and explicable with reference to preconceived 
criteria for similarity, but as itself given in experience. For the sceptic, a useful first step 
towards appreciating this view of what it is to know is to distinguish the EM model of a 
lift from an orthodox simplified formal model of a lift with a prescribed and preconceived 
interpretation and functionality. In the longer term, in the authorÕs opinion, the possibility 
of wider acceptance of RE and EM as a new framework for knowledge management does 
not rely upon such intellectual assent: it potentially offers such benefits in terms of the 
quality of the results and experience it can offer that its practical application will be 
justification in itself.
 
Acknowledgements: I am indebted to all the participants in the EM project for their 
various contributions and in particular to Cheryl Sidebotham for the basic EM lift model. 
I am especially indebted to Steve Russ for his role in promoting my interest in the 
philosophical aspects of EM and its application to issues of knowledge management, and 
for presenting an early version of this paper at the Workshop on Philosophy and 
Knowledge Management at Lucerne in April 2003. I also wish to thank Bertin Klein and 
Hans-Joachim Petsche for their help and encouragement.

References

1. W. M. Beynon, Empirical Modelling and the Foundations of AI, Computation for Metaphors, 
Analogy and Agents, LNAI 1562, Springer, 322-364, 1999
2. W. M. Beynon, EM for Educational Technology, Proc. Cognitive Technology '97, University 
of Aizu, Japan, IEEE, 54-68, 1997
3. W.M.Beynon et al, The Computer as Instrument, Proc. CT 2001, LNAI 2117, Springer 2001, 
476-489
4. W.M.Beynon, P-H Sun, Computer-mediated communication: a distributed Empirical 
Modelling perspective, Proc. CT 1999, San Francisco, August 1999, 115-132
5. W.M.Beynon, S.B.Russ, The development and use of variables in mathematics and computer 
science, The Mathematical Revolution Inspired by Computing, IMA Conf. Series 30, 25-95, 1991
6. W.M.Beynon, J.Rungrattanaubol, J.Sinclair, Formal Specification from an Observation-
Oriented Perspective, JUCS, Vol. 6 (4), 2000, 407-421
7. W.M.Beynon, S. Rasmequan, S.B.Russ, A new paradigm for computer-based decision support, 
Decision Support Systems, vol 33, no2, 2002, 127-142
8. G. Bird, William James, Routledge and Kegan Paul, 1986
9. B. Cantwell-Smith, The Foundations of Computing, 
http://www.ageofsig.org/people/bcsmith/print/smith-foundtns.pdf (accessed 7/8/03) 
10. M.L.Ginsberg,  D.E.Smith, Reasoning about Action I: A Possible Worlds Approach, Artificial 
Intelligence 35 (1988), 165-195
11. D. Gooding, Experiment and the making of meaning, Kluwer 1990
12. I. Jacobson, M. Christerson, P. Jonsson, G. Overgaard. Object-oriented Software Engineering: 
A Use-Case Driven Approach, Addison-Wesley 1992
13. William James, Essays in Radical Empiricism, Bison Books, 1996
14. C. Levi-Strauss, The Savage Mind, University of Chicago Press, 1966 
15. J. Rungrattanaubol, A Treatise on Modelling with Definitive Scripts, PhD Thesis, Computer 
Science Department, University of Warwick , April 2002
16. M. Turner, The Literary Mind, Oxford University Press, 1996
17. The EM website at URL http://www.dcs.warwick.ac.uk/modelling/
18. The EM archive at URL http://empublic.dcs.warwick.ac.uk/projects/
25