One of the challenges in multi-agent systems comes from the environmental non-stationarity that policies of all agents are evolving individually over time. Many existing multi-agent reinforcement learning (MARL) methods have been proposed to address this problem. However, these methods rely on a large amount of training data and some of them require agents to intensely communicate, which is often impractical in real-world applications. To better tackle the non-stationarity problem, this article combines model-based reinforcement learning (MBRL) and meta-learning and proposes a method called Model-based Imagined Rollouts Augmentation (MIRA). Based on an environment dynamics model, distributed agents can independently perform multi-agent rollouts with opponent models during exploitation and learn to infer the environmental non-stationarity as a latent variable using the rollouts. Based on the world model and latent-variable inference module, we perform multi-agent soft actor-critic implementation for centralized training and decentralized decision making. Empirical results on the Multi-agent Particle Environment (MPE) have proved that the algorithm has a very considerable improvement in sample efficiency as well as better convergent rewards than state-of-the-art MARL methods, including COMA, MAAC, MADDPG, and VDN.

When agents need to collaborate without previous coordination, the multi-agent cooperation problem transforms into an ad hoc teamwork (AHT) problem. Mainstream research on AHT is divided into type-based and type-free methods. The former depends on known teammate types to infer the current teammate type, while the latter does not require them at all. However, in many real-world applications, the complete absence and sufficient knowledge of known types are both impractical. Thus, this research focuses on the challenge of AHT with limited known types. To this end, this paper proposes a method called a Few typE-based Ad hoc Teamwork via meta-reinforcement learning (FEAT), which effectively adapts to teammates using a small set of known types within a single episode. FEAT enables agents to develop a highly adaptive policy through meta-reinforcement learning by employing limited priors about known types. It also utilizes this policy to generate a diverse type repository automatically. During the ad hoc cooperation, the agent can autonomously identify known teammate types followed by directly utilizing the pre-trained optimal cooperative policy or swiftly updating the meta policy to respond to teammates of unknown types. Comprehensive experiments in the pursuit domain validate the effectiveness of the algorithm and its components.