Deep Mind’s Alpha Go, the Go-playing engine that defeated the World Go Champion Lee Sedol in March 2016, is one of the biggest achievements of Artificial Intelligence.
In a nutshell, Alpha Go is a decision-making agent that takes decisions in an uncertain environment, exploring the potential consequences of their own choices using complex estimates of the world around.
This course is a study of the basic building blocks of decision-making agents, which are abstract entities living in an uncertain environment and are guided towards the realisation of given objectives.
An agent is typically endowed with a knowledge base, a collection of facts expressed in some logical language, and an action repertoire at each state. The agent can reason about the environment, using their knowledge base, and take decisions accordingly. The environment is typically unknown, stochastic, and evolves following some rules that might be unknown to the agent, as well. On top of this, it is usually inhabited by other agents, which may or may not strive to achieve similar objectives. The task is to take the best possible decision that can be taken given the (incomplete) information available.
This simple model is the basis of a number of important achievements in AI, and combines the use of logical, game-theoretic and algorithmic analysis.
The course will be an exploration of the basic methodologies for the design of artificial agents in complex environments. The course will first start with classical AI approaches where these agents are goal-oriented and take decisions in a potentially unknown environment.
Then it will move on to more sophisticated models allowing agents to have a representation of the other agents, their potential decisions and their goal, a representation about the representations of other agents, and so forth. This induces complex patterns of strategic reasoning, both in competitive and cooperative interactions, which need to be formally modelled and analysed.
These agent-based systems are built upon three important methodologies: Logic, because of the focus on reasoning, Game-Theory, because of the focus on strategies, and Algorithms, because of the focus on artificial agents.
You do not need any prior knowledge of these fields to study this course and it is open to anyone.
- Agents: definitions, applications
- Reasoning: logic and agents, knowledge representation, inference mechanisms
- Decision-making: actions, time and risk
- Learning: introduction to reinforcement learning
- Introduction to multi-agent systems: definitions, strategies and knowledge, collective strategies, agent application areas.
- Multi-agent reasoning: multi-agent epistemic logic, action logics, deliberation, BDI models.
- Modelling opponents: uncertainty and expectations, multi-agent learning.
- Competitive models: strategies and equilibria, opponent modelling.
- Cooperative models: bargaining and negotiation, resource allocation, inter-agent relationships.
- Open Issues: development methodology, programming languages, standards.
The course will be an investigation of the most important developments of AI in multi-agent contexts, touching upon themes such as opponent modelling, games with imperfect information, resource allocation, collective decision-making and electronic commerce applications.
Students will learn the basic methodologies for the design and the analysis of AI in complex systems with many interacting agents, ranging from competitive to cooperative interaction. The course take will be interdisciplinary, touching upon themes that are important for computer science, economics, and philosophy.
By the end of the course the students will learn how to programme a strategic agent participating in an auction. During the allocated seminars time students will be receive support on the programming skills required for the task (Python), from scratch.
The format will consist of three hours of lectures, followed by one hour of lab exercises, throughout the whole module. Seminar time will be devoted both to theory and to practice.
Exercise sheets will be provided, as well as samples of previous exam papers of mine.
A 2 hour exam (80%)and coursework (20%)
- Shoham Y. and Leyton-Brown K., Multi-Agent Systems: Logical, Algorithmic and Game Theoretic foundations, Cambridge University Press, 2009.
- Russell S and Norvig P, Artificial Intelligence: A Modern Approach, 3rd edition, Prentice-Hall, 2014.
- Shoham Y. and Leyton-Brown K., Essentials of Game Theory: A concise multidisciplinary introduction, Morgan & Claypool, 2008.
- Maschler M., Solan E. and Zamir S., Game Theory, Cambridge University Press, 2013.
- Wooldridge M., An Introduction to MultiAgent Systems, 2nd Edition, Wiley, 2009.
Materials will be updated here after each lecture.
Please note the details of the course content may be subject to change