Towards a Reinforcement Learning Module for Navigation in Video Games
Résumé
Large real-world Probabilistic Temporal Planning (PTP) is a very challenging research field. A common approach is to model such problems as Markov Decision Problems (MDP) and use dynamic programming techniques. Yet, two major difficulties arise: 1- dynamic programming does not scale with the number of tasks, and 2- the probabilistic model may be uncertain, leading to the choice of unsafe policies. We build here on the Factored Policy Gradient (FPG) algorithm and on robust decision-making to address both difficulties through an algorithm that trains two competing teams of learning agents. As the learning is simultaneous, each agent is facing a non-stationary environment. The goal is for them to find a common Nash equilibrium.