Offline multi-task reinforcement learning aims to learn a unified policy capable of solving multiple tasks using only pre-collected task-mixed datasets, without requiring any online interaction with the environment. However, it faces significant challenges in effectively sharing knowledge across tasks. Inspired by the efficient knowledge abstraction observed in human learning, we propose Goal-Oriented Skill Abstraction (GO-Skill), a novel approach designed to extract and utilize reusable skills to enhance knowledge transfer and task performance. Our approach uncovers reusable skills through a goal-oriented skill extraction process and leverages vector quantization to construct a discrete skill library. To mitigate class imbalances between broadly applicable and task-specific skills, we introduce a skill enhancement phase to refine the extracted skills. Furthermore, we integrate these skills using hierarchical policy learning, enabling the construction of a high-level policy that dynamically orchestrates discrete skills to accomplish specific tasks. Extensive experiments on diverse robotic manipulation tasks within the MetaWorld benchmark demonstrate the effectiveness and versatility of GO-Skill.

Multi-agent reinforcement learning (MARL) has demonstrated its superiority in addressing complex decision-making tasks involving multiple agents. However, the intricate and dynamic interactions among agents make this problem exceptionally challenging. Existing MARL methods often simplify the problem by implicitly decomposing shared rewards into individual utilities, neglecting the underlying interconnections between relevant entities. To overcome these limitations, we propose a novel framework of Multi-Agent Pattern Extraction (MAPE), which captures cooperation patterns from both agent-level and global perspectives to enhance decision-making and collaboration efficiency. Specifically, MAPE introduces two key modules: the Agent Pattern Extractor (APE) and the Global Pattern Extractor (GPE) that focus on specific interactions between the entities of interests from individual and global perspectives, respectively. The APE module focuses on computing each agent’s attention to other entities across different interaction patterns, providing this information to the GPE module. The GPE module then integrates the agent-specific pattern information and state features to identify the overall interaction pattern of the multi-agent system. By filtering out irrelevant interactions between unrelated entities and highlighting meaningful relationships, MAPE fosters more focused cooperation and facilitates more efficient learning. Extensive experiments on the StarCraft II micromanagement benchmark showcase the effectiveness of MAPE in improving efficiency in complex multi-agent environments.

Multi-Agent Reinforcement Learning (MARL) has achieved remarkable success in various real-world scenarios, but its high cost of online training makes it impractical to learn each task from scratch. To enable effective policy reuse, we consider the problem of zero-shot generalization from offline data across multiple tasks. While prior work focuses on transferring individual skills of agents, we argue that the effective policy transfer across tasks should also capture the team-level coordination knowledge. In this paper, we propose Bi-Level Knowledge Transfer (BiKT) for Multi-Task MARL, which performs knowledge transfer at both the individual and team levels. At the individual level, we extract transferable individual skill embeddings from offline MARL trajectories. At the team level, we define tactics as coordinated patterns of skill combinations and capture them by leveraging the learned skill embeddings. We map skill combinations into compact tactic embeddings and then construct a tactic codebook. To incorporate both skills and tactics into decision-making, we design a bi-level decision transformer that infers them in sequence. Our BiKT leverages both the generalizability of individual skills and the diversity of tactics, enabling the learned policy to perform effectively across multiple tasks. Extensive experiments on SMAC and MPE benchmarks demonstrate that BiKT achieves strong generalization to previously unseen tasks.

Multi-task reinforcement learning endeavors to accomplish a set of different tasks with a single policy. To enhance data efficiency by sharing parameters across multiple tasks, a common practice segments the network into distinct modules and trains a routing network to recombine these modules into task-specific policies. However, existing routing approaches employ a fixed number of modules for all tasks, neglecting that tasks with varying difficulties commonly require varying amounts of knowledge. This work presents a Dynamic Depth Routing (D2R) framework, which learns strategic skipping of certain intermediate modules, thereby flexibly choosing different numbers of modules for each task. Under this framework, we further introduce a ResRouting method to address the issue of disparate routing paths between behavior and target policies during off-policy training. In addition, we design an automatic route-balancing mechanism to encourage continued routing exploration for unmastered tasks without disturbing the routing of mastered ones. We conduct extensive experiments on various robotics manipulation tasks in the Meta-World benchmark, where D2R achieves state-of-the-art performance with significantly improved learning efficiency.

Efficient collaboration in the centralized training with decentralized execution (CTDE) paradigm remains a challenge in cooperative multi-agent systems. We identify divergent action tendencies among agents as a significant obstacle to CTDE's training efficiency, requiring a large number of training samples to achieve a unified consensus on agents' policies. This divergence stems from the lack of adequate team consensus-related guidance signals during credit assignment in CTDE. To address this, we propose Intrinsic Action Tendency Consistency, a novel approach for cooperative multi-agent reinforcement learning. It integrates intrinsic rewards, obtained through an action model, into a reward-additive CTDE (RA-CTDE) framework. We formulate an action model that enables surrounding agents to predict the central agent's action tendency. Leveraging these predictions, we compute a cooperative intrinsic reward that encourages agents to align their actions with their neighbors' predictions. We establish the equivalence between RA-CTDE and CTDE through theoretical analyses, demonstrating that CTDE's training process can be achieved using N individual targets. Building on this insight, we introduce a novel method to combine intrinsic rewards and RA-CTDE. Extensive experiments on challenging tasks in SMAC, MPE, and GRF benchmarks showcase the improved performance of our method.

Opponent modeling aims at learning the opponent's behaviors, goals, or beliefs to reduce the uncertainty of the competitive environment and assist decision-making. Existing work has mostly focused on learning opponent models online, which is impractical and inefficient in practical scenarios. To this end, we formalize an Offline Opponent Modeling (OOM) problem with the objective of utilizing pre-collected offline datasets to learn opponent models that characterize the opponent from the viewpoint of the controlled agent, which aids in adapting to the unknown fixed policies of the opponent. Drawing on the promises of the Transformers for decision-making, we introduce a general approach, Transformer Against Opponent (TAO), for OOM. Essentially, TAO tackles the problem by harnessing the full potential of the supervised pre-trained Transformers' in-context learning capabilities. The foundation of TAO lies in three stages: an innovative offline policy embedding learning stage, an offline opponent-aware response policy training stage, and a deployment stage for opponent adaptation with in-context learning. Theoretical analysis establishes TAO's equivalence to Bayesian posterior sampling in opponent modeling and guarantees TAO's convergence in opponent policy recognition. Extensive experiments and ablation studies on competitive environments with sparse and dense rewards demonstrate the impressive performance of TAO. Our approach manifests remarkable prowess for fast adaptation, especially in the face of unseen opponent policies, confirming its in-context learning potency.

Multi-task reinforcement learning endeavors to efficiently leverage shared information across various tasks, facilitating the simultaneous learning of multiple tasks. Existing approaches primarily focus on parameter sharing with carefully designed network structures or tailored optimization procedures. However, they overlook a direct and complementary way to exploit cross-task similarities: the control policies of tasks already proficient in some skills can provide explicit guidance for unmastered tasks to accelerate skills acquisition. To this end, we present a novel framework called Cross-Task Policy Guidance (CTPG), which trains a guide policy for each task to select the behavior policy interacting with the environment from all tasks' control policies, generating better training trajectories. In addition, we propose two gating mechanisms to improve the learning efficiency of CTPG: one gate filters out control policies that are not beneficial for guidance, while the other gate blocks tasks that do not necessitate guidance. CTPG is a general framework adaptable to existing parameter sharing approaches. Empirical evaluations demonstrate that incorporating CTPG with these approaches significantly enhances performance in manipulation and locomotion benchmarks.

Opponent modeling is a longstanding research topic aimed at enhancing decision-making by modeling information about opponents in multi-agent environments. However, existing approaches often face challenges such as having difficulty generalizing to unknown opponent policies and conducting unstable performance. To tackle these challenges, we propose a novel approach based on in-context learning and decision-time search named Opponent Modeling with In-context Search (OMIS). OMIS leverages in-context learning-based pretraining to train a Transformer model for decision-making. It consists of three in-context components: an actor learning best responses to opponent policies, an opponent imitator mimicking opponent actions, and a critic estimating state values. When testing in an environment that features unknown non-stationary opponent agents, OMIS uses pretrained in-context components for decision-time search to refine the actor's policy. Theoretically, we prove that under reasonable assumptions, OMIS without search converges in opponent policy recognition and has good generalization properties; with search, OMIS provides improvement guarantees, exhibiting performance stability. Empirically, in competitive, cooperative, and mixed environments, OMIS demonstrates more effective and stable adaptation to opponents than other approaches. See our project website at https://sites.google.com/view/nips2024-omis.

Despite the potential of Multi-Agent Reinforcement Learning (MARL) in addressing numerous complex tasks, training a single team of MARL agents to handle multiple diverse team tasks remains a challenge. In this paper, we introduce a novel Multi-task method based on Knowledge Transfer in cooperative MARL (MKT-MARL). By learning from task-specific teachers, our approach empowers a single team of agents to attain expert-level performance in multiple tasks. MKT-MARL utilizes a knowledge distillation algorithm specifically designed for the multi-agent architecture, which rapidly learns a team control policy incorporating common coordinated knowledge from the experience of task-specific teachers. Additionally, we enhance this training with teacher annealing, gradually shifting the model's learning from distillation towards environmental rewards. This enhancement helps the multi-task model surpass its single-task teachers. We extensively evaluate our algorithm using two commonly-used benchmarks: StarCraft II micro-management and multi-agent particle environment. The experimental results demonstrate that our algorithm outperforms both the single-task teachers and a jointly-trained team of agents. Extensive ablation experiments illustrate the effectiveness of the supervised knowledge transfer and the teacher annealing strategy.

Grouping is ubiquitous in natural systems and is essential for promoting efficiency in team coordination. This paper proposes a novel formulation of Group-oriented Multi-Agent Reinforcement Learning (GoMARL), which learns automatic grouping without domain knowledge for efficient cooperation. In contrast to existing approaches that attempt to directly learn the complex relationship between the joint action-values and individual utilities, we empower subgroups as a bridge to model the connection between small sets of agents and encourage cooperation among them, thereby improving the learning efficiency of the whole team. In particular, we factorize the joint action-values as a combination of group-wise values, which guide agents to improve their policies in a fine-grained fashion. We present an automatic grouping mechanism to generate dynamic groups and group action-values. We further introduce a hierarchical control for policy learning that drives the agents in the same group to specialize in similar policies and possess diverse strategies for various groups. Experiments on the StarCraft II micromanagement tasks and Google Research Football scenarios verify our method's effectiveness. Extensive component studies show how grouping works and enhances performance.

Opponent modeling is essential to exploit sub-optimal opponents in strategic interactions. Most previous works focus on building explicit models to predict the opponents' styles or strategies, which require a large amount of data to train the model and lack adaptability to unknown opponents. In this work, we propose a novel Learning to Exploit (L2E) framework for implicit opponent modeling. L2E acquires the ability to exploit opponents through a few interactions with different opponents during training of a neural network and can quickly adapt to new opponents with unknown styles during testing. To automatically produce challenging and diverse opponents for training, we further present a novel opponent strategy generation algorithm. We evaluate L2E on two poker games and one grid soccer game, which are the commonly used benchmarks for opponent modeling. Comprehensive experimental results indicate that L2E rapidly adapts to diverse styles of unknown opponents.

Experience replay plays a crucial role in Reinforcement Learning (RL), enabling the agent to remember and reuse experience from the past. Most previous methods sample experience transitions using simple heuristics like uniformly sampling or prioritizing those good ones. Since humans can learn from both good and bad experiences, more sophisticated experience replay algorithms need to be developed. Inspired by the potential energy in physics, this work introduces the artificial potential field into experience replay and develops Potentialized Experience Replay (PotER) as a new and effective sampling algorithm for RL in hard exploration tasks with sparse rewards. PotER defines a potential energy function for each state in experience replay and helps the agent to learn from both good and bad experiences using intrinsic state supervision. PotER can be combined with different RL algorithms as well as the self-imitation learning algorithm. Experimental analyses and comparisons on multiple challenging hard exploration environments have verified its effectiveness and efficiency.