This paper considers a novel multi-agent linear stochastic approximation algorithm driven by Markovian noise and general consensus-type interaction, in which each agent evolves according to its local stochastic approximation process which depends on the information from its neighbors. The interconnection structure among the agents is described by a time-varying directed graph. While the convergence of consensus-based stochastic approximation algorithms when the interconnection among the agents is described by doubly stochastic matrices (at least in expectation) has been studied, less is known about the case when the interconnection matrix is simply stochastic. For any uniformly strongly connected graph sequences whose associated interaction matrices are stochastic, the paper derives finite-time bounds on the mean-square error, defined as the deviation of the output of the algorithm from the unique equilibrium point of the associated ordinary differential equation. For the case of interconnection matrices being stochastic, the equilibrium point can be any unspecified convex combination of the local equilibria of all the agents in the absence of communication. Both the cases with constant and time-varying step-sizes are considered. In the case when the convex combination is required to be a straight average and interaction between any pair of neighboring agents may be uni-directional, so that doubly stochastic matrices cannot be implemented in a distributed manner, the paper proposes a push-sum-type distributed stochastic approximation algorithm and provides its finite-time bound for the time-varying step-size case by leveraging the analysis for the consensus-type algorithm with stochastic matrices and developing novel properties of the push-sum algorithm. Distributed temporal difference learning is discussed as an illustrative application.

In this paper, we propose a new approach called Adaptive Behavioral Costs in Reinforcement Learning (ABC-RL) for training a human-like agent with competitive strength. While deep reinforcement learning agents have recently achieved superhuman performance in various video games, some of these unconstrained agents may exhibit actions, such as shaking and spinning, that are not typically observed in human behavior, resulting in peculiar gameplay experiences. To behave like humans and retain similar performance, ABC-RL augments behavioral limitations as cost signals in reinforcement learning with dynamically adjusted weights. Unlike traditional constrained policy optimization, we propose a new formulation that minimizes the behavioral costs subject to a constraint of the value function. By leveraging the augmented Lagrangian, our approach is an approximation of the Lagrangian adjustment, which handles the trade-off between the performance and the human-like behavior. Through experiments conducted on 3D games in DMLab-30 and Unity ML-Agents Toolkit, we demonstrate that ABC-RL achieves the same performance level while significantly reducing instances of shaking and spinning. These findings underscore the effectiveness of our proposed approach in promoting more natural and human-like behavior during gameplay.

Enabling bipedal walking robots to learn how to maneuver over highly uneven, dynamically changing terrains is challenging due to the complexity of robot dynamics and interacted environments. Recent advancements in learning from demonstrations have shown promising results for robot learning in complex environments. While imitation learning of expert policies has been well-explored, the study of learning expert reward functions is largely under-explored in legged locomotion. This paper brings state-of-the-art Inverse Reinforcement Learning (IRL) techniques to solving bipedal locomotion problems over complex terrains. We propose algorithms for learning expert reward functions, and we subsequently analyze the learned functions. Through nonlinear function approximation, we uncover meaningful insights into the expert's locomotion strategies. Furthermore, we empirically demonstrate that training a bipedal locomotion policy with the inferred reward functions enhances its walking performance on unseen terrains, highlighting the adaptability offered by reward learning.

We present a new algorithm based on posterior sampling for learning in constrained Markov decision processes (CMDP) in the infinite-horizon undiscounted setting. The algorithm achieves near-optimal regret bounds while being advantageous empirically compared to the existing algorithms. Our main theoretical result is a Bayesian regret bound for each cost component of \tilde{O} (HS \sqrt{AT}) for any communicating CMDP with S states, A actions, and bound on the hitting time H. This regret bound matches the lower bound in order of time horizon T and is the best-known regret bound for communicating CMDPs in the infinite-horizon undiscounted setting. Empirical results show that, despite its simplicity, our posterior sampling algorithm outperforms the existing algorithms for constrained reinforcement learning.

In multi-task reinforcement learning, it is possible to improve the data efficiency of training agents by transferring knowledge from other different but related tasks. Because the experiences from different tasks are usually biased toward the specific task goals. Traditional methods rely on Kullback-Leibler regularization to stabilize the transfer of knowledge from one task to the others. In this work, we explore the direction of replacing the Kullback-Leibler divergence with a novel Optimal transport-based regularization. By using the Sinkhorn mapping, we can approximate the Optimal transport distance between the state distribution of tasks. The distance is then used as an amortized reward to regularize the amount of sharing information. We experiment our frameworks on several grid-based navigation multi-goal to validate the effectiveness of the approach. The results show that our added Optimal transport-based rewards are able to speed up the learning process of agents and outperforms several baselines on multi-task learning.

In order to assess the risks of a network system, it is important to investigate the behaviors of attackers after successful exploitation, which is called post-exploitation. Although there are various efficient tools supporting post-exploitation implementation, no application can automate this process. Most of the steps of this process are completed by experts who have profound knowledge of security, known as penetration testers or pen-testers. To this end, our study proposes the Raij\=u framework, a Reinforcement Learning (RL)-driven automation approach that assists pen-testers in quickly implementing the process of post-exploitation for security-level evaluation in network systems. We implement two RL algorithms, Advantage Actor-Critic (A2C) and Proximal Policy Optimization (PPO), to train specialized agents capable of making intelligent actions, which are Metasploit modules to automatically launch attacks of privileges escalation, gathering hashdump, and lateral movement. By leveraging RL, we aim to empower these agents with the ability to autonomously select and execute actions that can exploit vulnerabilities in target systems. This approach allows us to automate certain aspects of the penetration testing workflow, making it more efficient and responsive to emerging threats and vulnerabilities. The experiments are performed in four real environments with agents trained in thousands of episodes. The agents automatically select actions and launch attacks on the environments and achieve over 84\% of successful attacks with under 55 attack steps given. Moreover, the A2C algorithm has proved extremely effective in the selection of proper actions for automation of post-exploitation.

Shifting from traditional control strategies to Deep Reinforcement Learning (RL) for legged robots poses inherent challenges, especially when addressing real-world physical constraints during training. While high-fidelity simulations provide significant benefits, they often bypass these essential physical limitations. In this paper, we experiment with the Constrained Markov Decision Process (CMDP) framework instead of the conventional unconstrained RL for robotic applications. We perform a comparative study of different constrained policy optimization algorithms to identify suitable methods for practical implementation. Our robot experiments demonstrate the critical role of incorporating physical constraints, yielding successful sim-to-real transfers, and reducing operational errors on physical systems. The CMDP formulation streamlines the training process by separately handling constraints from rewards. Our findings underscore the potential of constrained RL for the effective development and deployment of learned controllers in robotics.

Reinforcement learning(RL) algorithms face the challenge of limited data efficiency, particularly when dealing with high-dimensional state spaces and large-scale problems. Most of RL methods often rely solely on state transition information within the same episode when updating the agent's Critic, which can lead to low data efficiency and sub-optimal training time consumption. Inspired by human-like analogical reasoning abilities, we introduce a novel mesh information propagation mechanism, termed the 'Imagination Mechanism (IM)', designed to significantly enhance the data efficiency of RL algorithms. Specifically, IM enables information generated by a single sample to be effectively broadcasted to different states across episodes, instead of simply transmitting in the same episode. This capability enhances the model's comprehension of state interdependencies and facilitates more efficient learning of limited sample information. To promote versatility, we extend the IM to function as a plug-and-play module that can be seamlessly and fluidly integrated into other widely adopted RL algorithms. Our experiments demonstrate that IM consistently boosts four mainstream SOTA RL algorithms, such as SAC, PPO, DDPG, and DQN, by a considerable margin, ultimately leading to superior performance than before across various tasks. For access to our code and data, please visit https://github.com/OuAzusaKou/imagination_mechanism

Visual Reinforcement Learning (RL) agents trained on limited views face significant challenges in generalizing their learned abilities to unseen views. This inherent difficulty is known as the problem of $\textit{view generalization}$. In this work, we systematically categorize this fundamental problem into four distinct and highly challenging scenarios that closely resemble real-world situations. Subsequently, we propose a straightforward yet effective approach to enable successful adaptation of visual $\textbf{Mo}$del-based policies for $\textbf{Vie}$w generalization ($\textbf{MoVie}$) during test time, without any need for explicit reward signals and any modification during training time. Our method demonstrates substantial advancements across all four scenarios encompassing a total of $\textbf{18}$ tasks sourced from DMControl, xArm, and Adroit, with a relative improvement of $\mathbf{33}$%, $\mathbf{86}$%, and $\mathbf{152}$% respectively. The superior results highlight the immense potential of our approach for real-world robotics applications. Videos are available at https://yangsizhe.github.io/MoVie/ .

One of the key behavioral characteristics used in neuroscience to determine whether the subject of study -- be it a rodent or a human -- exhibits model-based learning is effective adaptation to local changes in the environment, a particular form of adaptivity that is the focus of this work. In reinforcement learning, however, recent work has shown that modern deep model-based reinforcement-learning (MBRL) methods adapt poorly to local environment changes. An explanation for this mismatch is that MBRL methods are typically designed with sample-efficiency on a single task in mind and the requirements for effective adaptation are substantially higher, both in terms of the learned world model and the planning routine. One particularly challenging requirement is that the learned world model has to be sufficiently accurate throughout relevant parts of the state-space. This is challenging for deep-learning-based world models due to catastrophic forgetting. And while a replay buffer can mitigate the effects of catastrophic forgetting, the traditional first-in-first-out replay buffer precludes effective adaptation due to maintaining stale data. In this work, we show that a conceptually simple variation of this traditional replay buffer is able to overcome this limitation. By removing only samples from the buffer from the local neighbourhood of the newly observed samples, deep world models can be built that maintain their accuracy across the state-space, while also being able to effectively adapt to local changes in the reward function. We demonstrate this by applying our replay-buffer variation to a deep version of the classical Dyna method, as well as to recent methods such as PlaNet and DreamerV2, demonstrating that deep model-based methods can adapt effectively as well to local changes in the environment.