IFlowNets Extend Generative Samplers for Incomplete Information Games

This research introduces Information Flow Networks (IFlowNets), a significant extension of generative flow networks tailored for incomplete information games. While existing methods often combine reinforcement learning (RL) and counterfactual regret (CFR) for game theory applications, this work explores a novel generative sampling framework. The authors demonstrate that previous constraints for generative flow networks, valid in complete information games, are insufficient for learning valid player strategies and objectives in incomplete information scenarios. IFlowNets address these limitations, providing a generalized framework that ensures valid densities corresponding to player strategies. Preliminary results across three standard game environments show that IFlowNets perform comparably to or even surpass Outcome Sampling Monte Carlo Counterfactual Regret (OSMCCFR) and traditional RL-based methods in both performance and computational speed. This advancement opens new avenues for developing AI agents capable of strategic decision-making in complex, real-world environments where information is often partial.