[ Français ]

Workshop postponed to November 16-18, 2021

Agents behaviour in combinatorial game theory

Real-world optimization processes frequently involve agents whose utilities do not match those of the decision-maker. As an example from the field of economics, first-best prices that would align the interests of selfish users with those of the society may not be available, yielding mathematical programs that must explicity embed population behavior within their formulation. This fits the framework of MPECs (Mathematical Programs with Equilibrium Constraints) which involve, even in their simplest instances, non-convexity and non-differentiability. MPECs are generically NP-hard, and the mere existence of a solution is not assured. This raises several issues that have led to computational advances in the fields of non-differentiable optimization, exact and approximate algorithms, heuristics and meta-heuristics, etc. Actually, an MPEC can be viewed as a leader-follower game, where the follower may involve several non-cooperative players, either atomic or non-atomic, and whose behavior must be adequately assessed, for instance through reinforcement learning approaches.

The aim of the workshop is to survey recent advances in large-scale games endowed with combinatorial features, with an emphasis on learning player behavior (preferences, utilities), a process closely related to data science and machine learning. Indeed, the relationships between these disciplines, together with optimization, will be at the core of the workshop. A defining feature of the workshop is that, having been exposed to perspectives that are usually regarded as territories of separate research communities, participants will widen their knowledge of the field.

algorithms logo