Search for: All records

Coevolving teams of agents promises effective solutions for many coordination tasks such as search and rescue missions or deep ocean exploration. Good team performance in such domains generally relies on agents discovering complex joint policies, which is particularly difficult when the fitness functions are sparse (where many joint policies return the same or even zero fitness values). In this paper, we introduce Novelty Seeking Multiagent Evolutionary Reinforcement Learning (NS-MERL), which enables agents to more efficiently explore their joint strategy space. The key insight of NS-MERL is to promote good exploratory behaviors for individual agents using a dense, novelty-based fitness function. Though the overall team-level performance is still evaluated via a sparse fitness function, agents using NS-MERL more efficiently explore their joint action space and more readily discover good joint policies. Our results in complex coordination tasks show that teams of agents trained with NS-MERL perform significantly better than agents trained solely with task-specific fitnesses. 

In multiagent systems that require coordination, agents must learn diverse policies that enable them to achieve their individual and team objectives. Multiagent Quality-Diversity methods partially address this problem by filtering the joint space of policies to smaller sub-spaces that make the diversification of agent policies tractable. However, in teams of asymmetric agents (agents with different objectives and capabilities), the search for diversity is primarily driven by the need to find policies that will allow agents to assume complementary roles required to work together in teams. This work introduces Asymmetric Island Model (AIM), a multiagent framework that enables populations of asymmetric agents to learn diverse complementary policies that foster teamwork via dynamic population size allocation on a wide variety of team tasks. The key insight of AIM is that the competitive pressure arising from the distribution of policies on different team-wide tasks drives the agents to explore regions of the policy space that yield specializations that generalize across tasks. Simulation results on multiple variations of a remote habitat problem highlight the strength of AIM in discovering robust synergies that allow agents to operate near-optimally in response to the changing team composition and policies of other agents. 

Multi-robot teams have been shown to be effective in accomplishing complex tasks which require tight coordination among team members. In homogeneous systems, recent work has demonstrated that “stepping stone” rewards are an effective way to provide agents with feedback on potentially valuable actions even when the agent-to-agent coupling require- ments of an objective are not satisfied. In this work, we propose a new mechanism for inferring hypothetical partners in tightly-coupled, heterogeneous systems called Dirichlet-Multinomial Counterfactual Selection (DMCS). Using DMCS, we show that agents can learn to infer appropriate counterfactual partners to receive more informative stepping stone rewards by testing in a modified multi-rover exploration problem. We also show that DMCS outperforms a random partner selection baseline by over 40%, and we demonstrate how domain knowledge can be used to induce a prior to guide the agent learning process. Finally, we show that DMCS maintains superior performance for up to 15 distinct rover types compared to the performance of the baseline which degrades rapidly. 

The state-action space of an individual agent in a multiagent team fundamentally dictates how the individual interacts with the rest of the team. Thus, how an agent is defined in the context of its domain has a significant effect on team performance when learning to coordinate. In this work we explore the trade-offs associated with these design choices, for example, having fewer agents in the team that individually are able to process and act on a wider scope of information about the world versus a larger team of agents where each agent observes and acts in a more local region of the domain. We focus our study on a traffic management domain and highlight the trends in learning performance when applying different agent definitions.