When applying Reinforcement Learning (RL) algorithms to vision-based tasks, the significant variations between training and actual working environments pose a challenge to their generalization capability. While previous methods can enhance the generalization of the base RL algorithm, they are often limited to cases with minor changes between training and working environments. In this paper, we propose an effective auxiliary task called Jacobian Matrix Meets Masked Contrastive Learning (J-Mac), which aims to enhance the base RL algorithm's generalization capability even when there are significant changes between training and working environments. Specifically, we learn the correlations between visual states via transition dynamic learning. Meanwhile, on this basis, we eliminate task-irrelevant features from the representation of the visual state via masked contrastive learning. Extensive experiments demonstrate that our approach significantly improves the generalization of various base RL algorithms, outperforming other state-of-the-art methods across different vision-based benchmarks.

Offline-to-online reinforcement learning (RL) provides a promising solution to improving suboptimal offline pre-trained policies through online fine-tuning. However, one efficient method, unconstrained fine-tuning, often suffers from severe policy collapse due to excessive distribution shift. To ensure stability, existing methods retain offline constraints and employ additional techniques during fine-tuning, which hurts efficiency. In this work, we introduce a novel perspective: eliminating the policy collapse without imposing constraints. We observe that such policy collapse arises from the mismatch between unconstrained fine-tuning and the conventional RL training framework. To this end, we propose Stabilized Unconstrained Fine-tuning (SUF), a streamlined framework that benefits from the efficiency of unconstrained fine-tuning while ensuring stability by modifying the Update-To-Data ratio. With just a few lines of code adjustments, SUF demonstrates remarkable adaptability to diverse backbones and superior performance over state-of-the-art baselines.

We propose an end-to-end learning framework based on hierarchical reinforcement learning, called H-TSP, for addressing the large-scale Traveling Salesman Problem (TSP). The proposed H-TSP constructs a solution of a TSP instance starting from the scratch relying on two components: the upper-level policy chooses a small subset of nodes (up to 200 in our experiment) from all nodes that are to be traversed, while the lower-level policy takes the chosen nodes as input and outputs a tour connecting them to the existing partial route (initially only containing the depot). After jointly training the upper-level and lower-level policies, our approach can directly generate solutions for the given TSP instances without relying on any time-consuming search procedures. To demonstrate effectiveness of the proposed approach, we have conducted extensive experiments on randomly generated TSP instances with different numbers of nodes. We show that H-TSP can achieve comparable results (gap 3.42% vs. 7.32%) as SOTA search-based approaches, and more importantly, we reduce the time consumption up to two orders of magnitude (3.32s vs. 395.85s). To the best of our knowledge, H-TSP is the first end-to-end deep reinforcement learning approach that can scale to TSP instances of up to 10000 nodes. Although there are still gaps to SOTA results with respect to solution quality, we believe that H-TSP will be useful for practical applications, particularly those that are time-sensitive e.g., on-call routing and ride hailing service.

Proximal Policy Optimization (PPO) is a stable on-policy policy gradient (PG) method thanks to its clipped importance sampling (IS) weight objective of policy improvement. However, on-policy PG methods usually suffer from poor sample efficiency. In contrast, off-policy methods have demonstrated better sample efficiency by making more effective use of all collected samples during training. In this work, we aim to develop methods that inherit both the stability of on-policy PG methods and the data efficiency of off-policy methods. To this end, we present GeDISC, an off-policy algorithm that improves sample efficiency by reusing off-policy samples drawn from prior policies. Besides, we propose double IS weight clipping to control the high instability caused by off-policy data. We take the recently proposed generalized clipping mechanism for off-policy data as the first clipping to bound the policy update from the current policy and meanwhile we extend the standard clipping mechanism in PPO as the second clipping to prevent high variance and bias brought by extremely old samples. Extensive experiments on continuous and discrete control tasks show that the proposed new algorithm outperforms PPO and other SOTA PPO-based off-policy algorithms.

Recent approaches have utilized self-supervised auxiliary tasks as representation learning to improve the performance and sample efficiency of vision-based reinforcement learning algorithms in single-agent settings. However, in multi-agent reinforcement learning (MARL), these techniques face challenges because each agent only receives partial observation from an environment influenced by others, resulting in correlated observations in the agent dimension. So it is necessary to consider agent-level information in representation learning for MARL. In this paper, we propose an effective framework called Multi-Agent Masked Attentive Contrastive Learning (MA2CL), which encourages learning representation to be both temporal and agent-level predictive by reconstructing the masked agent observation in latent space. Specifically, we use an attention reconstruction model for recovering and the model is trained via contrastive learning. MA2CL allows better utilization of contextual information at the agent level, facilitating the training of MARL agents for cooperation tasks. Extensive experiments demonstrate that our method significantly improves the performance and sample efficiency of different MARL algorithms and outperforms other methods in various vision-based and state-based scenarios.

The increased integration of renewable energy poses a slew of technical challenges for the operation of power distribution networks. Among them, voltage fluctuations caused by the instability of renewable energy are receiving increasing attention. Utilizing MARL algorithms to coordinate multiple control units in the grid, which is able to handle rapid changes of power systems, has been widely studied in active voltage control task recently. However, existing approaches based on MARL ignore the unique nature of the grid and achieve limited performance. In this paper, we introduce the transformer architecture to extract representations adapting to power network problems and propose a Transformer-based Multi-Agent Actor-Critic framework (T-MAAC) to stabilize voltage in power distribution networks. In addition, we adopt a novel auxiliary-task training process tailored to the voltage control task, which improves the sample efficiency and facilitating the representation learning of the transformer-based model. We couple T-MAAC with different multi-agent actor-critic algorithms, and the consistent improvements on the active voltage control task demonstrate the effectiveness of the proposed method.

Learning good feature representations is important for deep reinforcement learning (RL). However, with limited experience, RL often suffers from data inefficiency for training. For un-experienced or less-experienced trajectories (i.e., state-action sequences), the lack of data limits the use of them for better feature learning. In this work, we propose a novel method, dubbed PlayVirtual, which augments cycle-consistent virtual trajectories to enhance the data efficiency for RL feature representation learning. Specifically, PlayVirtual predicts future states in a latent space based on the current state and action by a dynamics model and then predicts the previous states by a backward dynamics model, which forms a trajectory cycle. Based on this, we augment the actions to generate a large amount of virtual state-action trajectories. Being free of groudtruth state supervision, we enforce a trajectory to meet the cycle consistency constraint, which can significantly enhance the data efficiency. We validate the effectiveness of our designs on the Atari and DeepMind Control Suite benchmarks. Our method achieves the state-of-the-art performance on both benchmarks. Our code is available at https://github.com/microsoft/Playvirtual.