HSVI-based Online Minimax Strategies for Partially Observable Stochastic Games with Neural Perception Mechanisms