Coronavirus (Covid-19): Latest updates and information
Skip to main content Skip to navigation

Yibo Zhao

Machine Educable Engine on Noughts And Crosses

There are some essential elements missing from your abstract and description of your modelling study. Firstly, it should be made absolutely clear that you are investigating a specific artefact: Donald Michie's "Machine Educable Noughts And Crosses Engine (MENACE)" which was implemented using 300 matchboxes, each representing a unique board state ('MENACE' should appear explicitly in the title, and its original source should appear in your references). Secondly, you need to explain how you intend to approach the task of modelling MENACE using EM principles. Thirdly, you need to set out the reasons why this is an interesting study from an EM perspective, what you see as the main issues to be addressed and what might be learnt from your study. Having learnt a little more about this modelling study myself since I first proposed it, I now see it that your agenda is more difficult than I had imagined! I certainly don't think you will need to look beyond this particular exercise in order to find additional material and can focus your paper and model on the simulation of MENACE alone.

You have done useful scholarship to identify highly relevant resources, such as the specification of the AI exercise: Implementing MENACE at Uppsala University, Sweden. That resource gives useful background and insights that you can draw on e.g. the observations that "there are no resource conflicts - each box will only be used by one of the players" and "If the machine had played against an optimal opponent already from the start, it would not have worked". It also observes that the number of possible distinct board positions to be considered (for which 3^9 = 19683 is a crude upper bound) is restricted to approximately 300 "due to symmetries, rotations and since games end immediately after a win", but that such optimisation is probably unnecessary with a modern computer simulation. In the spirit of EM, the emphasis should be more visual than abstract algorithmic, and the same optimisation that led Michie to his 300 matchboxes should be applied in your context.

Bear in mind that a good solution to the Uppsala AI exercise is not necessarily going to earn high marks as a study in EM. Merely using EDEN "to build a machine educable engine in noughts and crosses" as you put it is not in itself evidence of engagement with EM thinking. When you tentatively suggest that this project "will use ‘tkeden’, because the algorithm overweight the implying environment", which I take to mean "because the algorithm outweighs the implementation environment in importance" I have two concerns. The implementation environment does matter in so far as you should be exploiting the scope for expressing key dependencies as EDEN definitions, and it isn't the 'algorithm' itself that is the primary target for implementation, it is the observables, dependencies and agency that surrounds the manipulation of matchboxes and their contents that first needs to be considered (think of the heapsort construal or more topically the OXO construals). It may not be possible to complete this modelling in detail within the scope of a WEB-EM submission, but you will get credit for what you are able to do so long as it is appropriately oriented.

The Uppsala students will be rewarded according to how effectively they implement a machine learning program without major concern for how this implementation is done. What we're looking for in EM is evidence that you have explored the challenge of modelling the observables, dependencies and agency in Michie's MENACE model. If you do this well - or even partially succeed in this - it should help both you and me to understand the nature and mechanisms of Michie's model much more clearly. At the moment, what you've written doesn't suggest good understanding of your topic - for instance, it is hardly likely that the technique that Michie is using here can scale up to positions in Pjawns even allowing for optimisations of the kind discussed above.

Writing an abstract and paper is definitely challenging as far as use of English is concerned. If you can make a good johb of the model-building I would weight your study as 70:30 for model:paper. Good to get a native speaker to read your final submission.

imply → implement throughout your text