Argumentation-Based Reinforcement Learning for RoboCup Soccer Takeaway

Reinforcement Learning (RL) is widely regarded as a generic and effective technique to learn coordinated behaviours in cooperative multi-agent systems (CMAS), but it suffers from slow convergence speed due to the huge joint action space. Incorporating domain knowledge has shown to be an effective method to tackle this problem, but little research has investigated how to propose high-quality heuristics. We consider a widely used CMAS application, RoboCup Takeaway, and use value-based argumentation to extract heuristics from conflicting domain knowledge therein.