We study the problem of designing an AI agent that can robustlycooperate with agents of unknown type (i.e., previously unobservedbehavior) in multi-agent scenarios. Our work is inspired by real-world applications in which an AI agent, e.g., a virtual assistant, hasto cooperate with new types of agents/users after its deployment.We model this problem via parametric Markov Decision Processeswhere the parameters correspond to a user’s type and characterizeher behavior. In the test phase, the AI agent has to interact with auser of an unknown type. We develop an algorithmic frameworkfor learning adaptive policies: our approach relies on observing theuser’s actions to make inferences about the user’s type and adaptingthe policy to facilitate efficient cooperation. We show that withoutbeing adaptive, an AI agent can end up performing arbitrarily badin the test phase. Using our framework, we propose two concretealgorithms for computing policies that automatically adapt to theuser in the test phase. We demonstrate the effectiveness of ouralgorithms in a cooperative gathering game environment for twoagents.KEYWORDSLearning agent-to-agent interactions; Machine learning; Reinforce-ment learning

Top