Rethinking formal models of partially observable multiagent decision making

Multiagent decision-making in partially observable environments is usually modelled as either an extensive-form game (EFG) theory or a stochastic (POSG) multiagent reinforcement learning (MARL). One issue with the current situation that while most practical problems can be both formalisms, relationship of two models unclear, which hinders transfer ideas between communities. A second EFGs have recently seen significant algorithmic progress, their classical formalization unsuitable for efficient presentation underlying ideas, such those around decomposition. To solve first issue, we introduce factored-observation games (FOSGs), minor modification POSG formalism distinguishes private and public observation thereby greatly simplifies remedy show FOSGs POSGs are naturally connected to EFGs: by “unrolling” FOSG into its tree form, obtain EFG. Conversely, any perfect-recall timeable EFG corresponds some this manner. Moreover, justifies several modifications appeared implicit response model's issues Finally, illustrate MARL presenting three key techniques – counterfactual regret minimization, sequence decomposition framework.