Many applications of imitation learning require the agent to generate the full distribution of behaviour observed in the training data. For example, to evaluate the safety of autonomous vehicles in simulation, accurate and diverse behaviour models of other road users are paramount. Existing methods that improve this distributional realism typically rely on hierarchical policies. These condition the policy on types such as goals or personas that give rise to multi-modal behaviour. However, such methods are often inappropriate for stochastic environments where the agent must also react to external factors: because agent types are inferred from the observed future trajectory during training, these environments require that the contributions of internal and external factors to the agent behaviour are disentangled and only internal factors, i.e., those under the agent's control, are encoded in the type. Encoding future information about external factors leads to inappropriate agent reactions during testing, when the future is unknown and types must be drawn independently from the actual future. We formalize this challenge as distribution shift in the conditional distribution of agent types under environmental stochasticity. We propose Robust Type Conditioning (RTC), which eliminates this shift with adversarial training under randomly sampled types. Experiments on two domains, including the large-scale Waymo Open Motion Dataset, show improved distributional realism while maintaining or improving task performance compared to state-of-the-art baselines.

VDN and QMIX are two popular value-based algorithms for cooperative MARL that learn a centralized action value function as a monotonic mixing of per-agent utilities. While this enables easy decentralization of the learned policy, the restricted joint action value function can prevent them from solving tasks that require significant coordination between agents at a given timestep. We show that this problem can be overcome by improving the joint exploration of all agents during training. Specifically, we propose a novel MARL approach called Universal Value Exploration (UneVEn) that learns a set of related tasks simultaneously with a linear decomposition of universal successor features. With the policies of already solved related tasks, the joint exploration process of all agents can be improved to help them achieve better coordination. Empirical results on a set of exploration games, challenging cooperative predator-prey tasks requiring significant coordination among agents, and StarCraft II micromanagement benchmarks show that UneVEn can solve tasks where other state-of-the-art MARL methods fail.

Decentralized MDPs (Dec-MDPs) provide a rigorous framework for collaborative multi-agent sequential decisionmaking under uncertainty. However, their computational complexity limits the practical impact. To address this, we focus on a class of Dec-MDPs consisting of independent collaborating agents that are tied together through a global reward function that depends upon their entire histories of states and actions to accomplish joint tasks. To overcome scalability barrier, our main contributions are: (a) We propose a new actor-critic based Reinforcement Learning (RL) approach for event-based Dec-MDPs using successor features (SF) which is a value function representation that decouples the dynamics of the environment from the rewards; (b) We then present Dec-ESR (Decentralized Event based Successor Representation) which generalizes learning for event-based Dec-MDPs using SF within an end-to-end deep RL framework; (c) We also show that Dec-ESR allows useful transfer of information on related but different tasks, hence bootstraps the learning for faster convergence on new tasks; (d) For validation purposes, we test our approach on a large multi-agent coverage problem which models schedule coordination of agents in a real urban subway network and achieves better quality solutions than previous best approaches.

Decentralized (PO)MDPs provide a rigorous framework for sequential multiagent decision making under uncertainty. However, their high computational complexity limits the practical impact. To address scalability and real-world impact, we focus on settings where a large number of agents primarily interact through complex joint-rewards that depend on their entire histories of states and actions. Such history-based rewards encapsulate the notion of events or tasks such that the team reward is given only when the joint-task is completed. Algorithmically, we contribute---1) A nonlinear programming (NLP) formulation for such event-based planning model; 2) A probabilistic inference based approach that scales much better than NLP solvers for a large number of agents; 3) A policy gradient based multiagent reinforcement learning approach that scales well even for exponential state-spaces. Our inference and RL-based advances enable us to solve a large real-world multiagent coverage problem modeling schedule coordination of agents in a real urban subway network where other approaches fail to scale.