Swarm intelligence in cooperative environments: N-step dynamic tree search algorithm extended analysis

Reinforcement learning tree-based planning methods have been gaining popularity in the last few years due to their success in single-agent domains, where a perfect simulator model is available, e.g., Go and chess strategic board games. This paper pretends to extend tree search algorithms to the multi-agent setting in a decentralized structure, dealing with scalability issues and exponential growth of computational resources. The N-Step Dynamic Tree Search combines forward planning and direct temporal-difference updates, outperforming markedly state-of-the-art algorithms such as Q-Learning and SARSA. Future state transitions and rewards are predicted with a model built and learned from real interactions between agents and the environment. As an extension of previous work, this paper analyses the developed algorithm in the Hunter-Pursuit cooperative game against intelligent evaders. The N-Step Dynamic Tree Search aims to adapt the most successful single-agent learning methods to the multi-agent boundaries and demonstrates to be a remarkable advance compared to conventional temporal-difference techniques.