Learning coordination in multi-agent systems
David Simões · Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT) · 2020
The ability for an agent to coordinate with others within a system is a valuable property in multi-agent systems. Agents either cooperate as a team to accomplish a common goal, or adapt to opponents to complete different goals without being exploited. Research has shown that learning multi-agent coordination is significantly more complex than learning policies in singleagent environments, and requires a variety of techniques to deal with the properties of a system where agents learn concurrently. This thesis aims to determine how can machine learning be used to achieve coordination within a multi-agent system. It asks what techniques can be used to tackle the increased complexity of such systems and their credit assignment challenges, how to achieve coordination, and how to use communication to improve the behavior of a team. Many algorithms for competitive environments are tabular-based, preventing their use with high-dimension or continuous state-spaces, and may be biased against specific equilibrium strategies. This thesis proposes multiple deep learning extensions for competitive environments, allowing algorithms to reach equilibrium strategies in complex and partially-observable environments, relying only on local information. A tabular algorithm is also extended with a new update rule that eliminates its bias against deterministic strategies. Current state-of-the-art approaches for cooperative environments rely on deep learning to handle the environment’s complexity and benefit from a centralized learning phase. Solutions that incorporate communication between agents often prevent agents from being executed in a distributed manner. This thesis proposes a multi-agent algorithm where agents learn communication protocols to compensate for local partial-observability, and remain independently executed. A centralized learning phase can incorporate additional environment information to increase the robustness and speed with which a team converges to successful policies. The algorithm outperforms current state-of-the-art approaches in a wide variety of multi-agent environments. A permutation invariant network architecture is also proposed to increase the scalability of the algorithm to large team sizes. Further research is needed to identify how can the techniques proposed in this thesis, for cooperative and competitive environments, be used in unison for mixed environments, and whether they are adequate for general artificial intelligence.