Centralized Training for Decentralized Execution, where agents are trained offline using centralized information but execute in a decentralized manner online, has gained popularity in the multi-agent reinforcement learning community. In particular, actor-critic methods with a centralized critic and decentralized actors are a common instance of this idea. However, the implications of using a centralized critic in this context are not fully discussed and understood even though it is the standard choice of many algorithms. We therefore formally analyze centralized and decentralized critic approaches, providing a deeper understanding of the implications of critic choice. Because our theory makes unrealistic assumptions, we also empirically compare the centralized and decentralized critic methods over a wide set of environments to validate our theories and to provide practical advice. We show that there exist misconceptions regarding centralized critics in the current literature and show that the centralized critic design is not strictly beneficial, but rather both centralized and decentralized critics have different pros and cons that should be taken into account by algorithm designers.

Recent advances in multi-agent reinforcement learning have been largely limited in training one model from scratch for every new task. The limitation is due to the restricted model architecture related to fixed input and output dimensions. This hinders the experience accumulation and transfer of the learned agent over tasks with diverse levels of difficulty (e.g. 3 vs 3 or 5 vs 6 multi-agent games). In this paper, we make the first attempt to explore a universal multi-agent reinforcement learning pipeline, designing one single architecture to fit tasks with the requirement of different observation and action configurations. Unlike previous RNN-based models, we utilize a transformer-based model to generate a flexible policy by decoupling the policy distribution from the intertwined input observation with an importance weight measured by the merits of the self-attention mechanism. Compared to a standard transformer block, the proposed model, named as Universal Policy Decoupling Transformer (UPDeT), further relaxes the action restriction and makes the multi-agent task's decision process more explainable. UPDeT is general enough to be plugged into any multi-agent reinforcement learning pipeline and equip them with strong generalization abilities that enables the handling of multiple tasks at a time. Extensive experiments on large-scale SMAC multi-agent competitive games demonstrate that the proposed UPDeT-based multi-agent reinforcement learning achieves significant results relative to state-of-the-art approaches, demonstrating advantageous transfer capability in terms of both performance and training speed (10 times faster).

In recent years, multi-agent deep reinforcement learning has been successfully applied to various complicated scenarios such as computer games and robot swarms. We thoroughly study and compare the state-of-the-art cooperative multi-agent deep reinforcement learning algorithms. Specifically, we investigate the consequences of the "hyperparameter tricks" of QMIX and its improved variants. Our results show that: (1) The significant performance improvements of these variant algorithms come from hyperparameter-level optimizations in their open-source codes (2) After modest tuning and with no changes to the network architecture, QMIX can attain extraordinarily high win rates in all hard and super hard scenarios of StarCraft Multi-Agent Challenge (SMAC) and achieve state-of-the-art (SOTA). In this work, we proposed a reliable QMIX benchmark, which will be of great benefit to subsequent research. Besides, we proposed a hypothesis to explain the excellent performance of QMIX.

Can artificial agents benefit from human conventions? Human societies manage to successfully self-organize and resolve the tragedy of the commons in common-pool resources, in spite of the bleak prediction of non-cooperative game theory. On top of that, real-world problems are inherently large-scale and of low observability. One key concept that facilitates human coordination in such settings is the use of conventions. Inspired by human behavior, we investigate the learning dynamics and emergence of temporal conventions, focusing on common-pool resources. Extra emphasis was given in designing a realistic evaluation setting: (a) environment dynamics are modeled on real-world fisheries, (b) we assume decentralized learning, where agents can observe only their own history, and (c) we run large-scale simulations (up to 64 agents). Uncoupled policies and low observability make cooperation hard to achieve; as the number of agents grow, the probability of taking a correct gradient direction decreases exponentially. By introducing an arbitrary common signal (e.g., date, time, or any periodic set of numbers) as a means to couple the learning process, we show that temporal conventions can emerge and agents reach sustainable harvesting strategies. The introduction of the signal consistently improves the social welfare (by 258% on average, up to 3306%), the range of environmental parameters where sustainability can be achieved (by 46% on average, up to 300%), and the convergence speed in low abundance settings (by 13% on average, up to 53%).

Conventional multi-agent path planners typically determine a path that optimizes a single objective, such as path length. Many applications, however, may require multiple objectives, say time-to-completion and fuel use, to be simultaneously optimized in the planning process. Often, these criteria may not be readily compared and sometimes lie in competition with each other. Simply applying standard multi-objective search algorithms to multi-agent path finding may prove to be inefficient because the size of the space of possible solutions, i.e., the Pareto-optimal set, can grow exponentially with the number of agents (the dimension of the search space). This paper presents an approach that bypasses this so-called curse of dimensionality by leveraging our prior multi-agent work with a framework called subdimensional expansion. One example of subdimensional expansion, when applied to A*, is called M* and M* was limited to a single objective function. We combine principles of dominance and subdimensional expansion to create a new algorithm named multi-objective M* (MOM*), which dynamically couples agents for planning only when those agents have to "interact" with each other. MOM* computes the complete Pareto-optimal set for multiple agents efficiently and naturally trades off sub-optimal approximations of the Pareto-optimal set and computational efficiency. Our approach is able to find the complete Pareto-optimal set for problem instances with hundreds of solutions which the standard multi-objective A* algorithms could not find within a bounded time.

Information theoretic sensor management approaches are an ideal solution to state estimation problems when considering the optimal control of multi-agent systems, however they are too computationally intensive for large state spaces, especially when considering the limited computational resources typical of large-scale distributed multi-agent systems. Reinforcement learning (RL) is a promising alternative which can find approximate solutions to distributed optimal control problems that take into account the resource constraints inherent in many systems of distributed agents. However, the RL training can be prohibitively inefficient, especially in low-information environments where agents receive little to no feedback in large portions of the state space. We propose a hybrid information-driven multi-agent reinforcement learning (MARL) approach that utilizes information theoretic models as heuristics to help the agents navigate large sparse state spaces, coupled with information based rewards in an RL framework to learn higher-level policies. This paper presents our ongoing work towards this objective. Our preliminary findings show that such an approach can result in a system of agents that are approximately three orders of magnitude more efficient at exploring a sparse state space than naive baseline metrics. While the work is still in its early stages, it provides a promising direction for future research.

When developing reinforcement learning agents, the standard approach is to train an agent to converge to a fixed policy that is as close to optimal as possible for a single fixed reward function. If different agent behaviour is required in the future, an agent trained in this way must normally be either fully or partially retrained, wasting valuable time and resources. In this study, we leverage multi-objective reinforcement learning to create tunable agents, i.e. agents that can adopt a range of different behaviours according to the designer's preferences, without the need for retraining. We apply this technique to sequential social dilemmas, settings where there is inherent tension between individual and collective rationality. Learning a single fixed policy in such settings leaves one at a significant disadvantage if the opponents' strategies change after learning is complete. In our work, we demonstrate empirically that the tunable agents framework allows easy adaption between cooperative and competitive behaviours in sequential social dilemmas without the need for retraining, allowing a single trained agent model to be adjusted to cater for a wide range of behaviours and opponent strategies.

Cooperative multi-agent reinforcement learning (MARL) has achieved significant results, most notably by leveraging the representation learning abilities of deep neural networks. However, large centralized approaches quickly become infeasible as the number of agents scale, and fully decentralized approaches can miss important opportunities for information sharing and coordination. Furthermore, not all agents are equal - in some cases, individual agents may not even have the ability to send communication to other agents or explicitly model other agents. This paper considers the case where there is a single, powerful, central agent that can observe the entire observation space, and there are multiple, low powered, local agents that can only receive local observations and cannot communicate with each other. The job of the central agent is to learn what message to send to different local agents, based on the global observations, not by centrally solving the entire problem and sending action commands, but by determining what additional information an individual agent should receive so that it can make a better decision. After explaining our MARL algorithm, hammer, and where it would be most applicable, we implement it in the cooperative navigation and multi-agent walker domains. Empirical results show that 1) learned communication does indeed improve system performance, 2) results generalize to multiple numbers of agents, and 3) results generalize to different reward structures.

Recent years saw a plethora of work on explaining complex intelligent agents. One example is the development of several algorithms that generate saliency maps which show how much each pixel attributed to the agents' decision. However, most evaluations of such saliency maps focus on image classification tasks. As far as we know, there is no work which thoroughly compares different saliency maps for Deep Reinforcement Learning agents. This paper compares four perturbation-based approaches to create saliency maps for Deep Reinforcement Learning agents trained on four different Atari 2600 games. All four approaches work by perturbing parts of the input and measuring how much this affects the agent's output. The approaches are compared using three computational metrics: dependence on the learned parameters of the agent (sanity checks), faithfulness to the agent's reasoning (input degradation), and run-time.

Training a multi-agent reinforcement learning (MARL) algorithm is more challenging than training a single-agent reinforcement learning algorithm, because the result of a multi-agent task strongly depends on the complex interactions among agents and their interactions with a stochastic and dynamic environment. We propose an algorithm that boosts MARL training using the biased action information of other agents based on a friend-or-foe concept. For a cooperative and competitive environment, there are generally two groups of agents: cooperative-agents and competitive-agents. In the proposed algorithm, each agent updates its value function using its own action and the biased action information of other agents in the two groups. The biased joint action of cooperative agents is computed as the sum of their actual joint action and the imaginary cooperative joint action, by assuming all the cooperative agents jointly maximize the target agent's value function. The biased joint action of competitive agents can be computed similarly. Each agent then updates its own value function using the biased action information, resulting in a biased value function and corresponding biased policy. Subsequently, the biased policy of each agent is inevitably subjected to recommend an action to cooperate and compete with other agents, thereby introducing more active interactions among agents and enhancing the MARL policy learning. We empirically demonstrate that our algorithm outperforms existing algorithms in various mixed cooperative-competitive environments. Furthermore, the introduced biases gradually decrease as the training proceeds and the correction based on the imaginary assumption vanishes.