Curiosity in Multi-Agent Reinforcement Learning
- Source: MSc thesis project at Autonomous Agents Research Group under supervision of Dr. Stefano Albrecht
- Type: Individual work
- Language(s): Python
In my MSc dissertation, I applied curiosity as intrinsically computed exploration bonuses to multi-agent reinforcement learning (MARL). Count- and prediction-based curiosity approaches were considered in combination with value-based and policy-gradient MARL methods, implemented using PyTorch. All approaches were evaluated using various competitive and cooperative MARL tasks in the multi-agent particle environment, also considering partial observability and sparse rewards, to analyse the influence of curiosity under such conditions. We found that curiosity led to considerably improved stability and convergence of policy-gradient MARL trained with sparse reward signals.