Multi-Agent Reinforcement Learning Design
Using automated design to promote multi-agent collaboration
A new focus of my work is to study the automated design of environments with multiple self-interested, reinforcement learning (RL) agents that share a set of limited resources. My objective is to use automated methods to understand why specific behaviors emerge in such settings, and to find the best way to design the environment in order to promote sustainable and socially-aware behaviors.
Many multi-agent systems maximize some performance criteria as if they were a single agent when in fact they are a collection of self-interested reinforcement learning agents. Examples include a company or firm, a group of robots in an assembly line that need to deliver items, a single robot with different components and more.
Dictating the behavior of agents in such settings via a centralized controller may either be impossible or at least impractical. We seek mechanisms that bridge the gap between the locally-desirable goals of individual agents and globally-desirable objectives. Market-based approaches replace the need for centralized control by allowing RL agents to trade the right to act in the environment.
The domain we are working on are multi-robot domains (e.g., robots can be rewarded for socially aware behaviors, such as removing obstacles from pathways even when these do not obstruct their own path) and sequential social dilemmas, which are games that model the tension between individual and global interests and are realistic model of many social situations.
Designing effective training environments for RL agents (Keren, Vashishtha, and Parkes 2019)
Interpretability of multi-agent RL solutions.