Adaptive Heuristics

Context

A single decision-making agent acting rationally in a stationary environment has to maximize its total expected reward. This is an optimization problem which can be solved using reinforcement learning methods like Q-learning.

In contrast, if decision-making agents interact strategically, i.e. the reward (or payoff) of one agent not only depends on its own decision but also of the decisions made by the others or in other words the agents are players in a game, the notion of rationality becomes much harder to define and solution concepts as Nash equilibrium and correlated equilibrium have been proposed to capture rationality in the context of strategic interaction.

If the agents are playing repeatedly the same game then learning based on past experience becomes a possibility.

Two important research questions related to learning in repeated games are

  • What is learning in the context of strategic interaction, and
  • Do these learning rules converge towards one of the solution concepts mentioned above?

A good starting point regarding these questions are the articles by Sergiu Hart (2005): "Adaptive Heuristics" and H. Peyton Young (2009) "Learning by Trial and Error".

Prerequisites

  • You must have taken (or, be willing to take) the 'Multi-Agent Learning' course.
  • You should have a working knowledge of English.

Research Training and Thesis

This will involve:

  • Make an overview and classify the different existing learning rules.
  • Implement known techniques and evaluate by doing computer experiments.
  • Analyze and compare results.

Contact

Interested students should contact Bernard Manderick (bmanderi@vub.ac.be) for more information.

Promotor: Bernard Manderick