Multi-agent deep reinforcement learning (MADRL) is a powerful approach that enables agents to coordinate and collaborate to solve complex collaborative tasks. Applications of MADRL range from robotics to traffic control and resource allocation. However, the learning process can be slow, especially with multiple agents, leading to low sample efficiency. To improve sample efficiency, one approach involves using adversaries to guide agents' exploration during the learning process. Adversaries act as perturbations, forcing agents to explore specific scenarios efficiently. When trained jointly, agents and adversaries can adapt their policies to each other, leading to promising results in challenging single-agent scenarios.
The objective of the thesis is to explore the use of adversaries to speed up the learning process within multi-agent settings. Different types of adversaries and their impact on the learning process will be investigated, and the effectiveness of the approach will be evaluated on multiple benchmark environments.