Knowledge Transfer for Deep Reinforcement Agents in General Game Playing

Learning to master new games with nothing but the rules given is a hallmark of human intelligence. This ability has recently been successfully replicated in AI systems through a combination of Knowledge Representation, Monte Carlo Tree Search and Deep Reinforcement Learning: Generalised AlphaZero [7] provides a method for building general game-playing agents that can learn any game describable in a formal specification language. We investigate how to boost the ability of deep reinforcement agents for general game playing by applying transfer learning for new game variants. Experiments show that transfer learning can significantly reduce the training time on variations of games that were previously learned, and our results further suggest that the most successful method is to train a source network that uses the guidance of multiple expert networks.