Neural Symbolic AI For POMDP Games

This paper demonstrates the difficulties of solving partially observable Markov decision process (POMDP) games with pure deep reinforcement learning. And shows that combining the strengths of deep reinforcement learning and symbolic reasoning can achieve improved results with less training time. Deep reinforcement learning excels at exploring states and optimizing actions based on those observations, but struggle to reason when facing uncertainty. So we propose an agent to use deep reinforcement learning to force exploration on the game states and its corresponding values, and then use reasoning on the given observations to deduce winning actions when arriving at high-value states, to enhance the performance compared to a pure deep reinforcement learning agent.