Counterfactual Regret Minimization in Anti-Submarine Warfare

Anti-Submarine Warfare (ASW) presents complex tactical decision challenges involving stealth and uncertainty. This project explores the application of Counterfactual Regret Minimization (CFR), a class of algorithms successful in large imperfect-information games, to ASW scenarios. We model the interaction between a submarine and an ASW force as a zero-sum, sequential game on discrete graphs with imperfect information, incorporating movement constraints, costs, and probabilistic detection. Using the OpenSpiel framework, we evaluate the performance of three CFR variants: vanilla CFR, Monte Carlo CFR (MCCFR), and Deep CFR, by comparing their convergence speed, measured in wall-clock time, to a target exploitability level. Experiments conducted on graphs of varying sizes showed that MCCFR converged the fastest on all graph sizes, while vanilla CFR was slightly slower; both significantly outperformed Deep CFR. Deep CFR, despite its theoretical scalability, exhibited significantly slower convergence, which may be attributed to computational overhead and hyperparameter sensitivity within the tested game's complexity range. These findings might suggest that for moderately sized strategic simulations like the ASW game modeled here, simpler CFR methods can be more computationally efficient than deep learning approaches.