Reactive Multi-Fitness Learning for Robust Multiagent Teaming
Connor Yates, Ayhan Alp Aydeniz, Kagan Tumer · 2021
Multi-robot systems deployed in remote missions for longitudinal tasks (tasks that occur over extended periods of time) are not only required to perform their initial tasks, but also need to react to changes in environmental conditions. In addition, performance in longitudinal missions is often difficult to capture in a single reward function. This work introduces Reactive Multi-Fitness Learning (R-MFL) to address both problems by separating “behaviors” and mission goals. R-MFL builds on Multi-Fitness Learning (MFL) that addresses which fitness matters when in complex and sequential tasks. R-MFL enables agents to change their set of behaviors instead of incorporating new insights into a single policy, while also determining when new behaviors are necessary. We show that the agents using R-MFL are able to react to unforeseen changes without catastrophic forgetting, and identify when and where they need to react during a deployment. R-MFL is shown to provide up to 90% improvement over MFL and 100% over a one-step evolutionary approach.