Multi-agent reinforcement learning (MARL) has found various applications in the field of transportation and simulating [50, 1], stock price analyzing and trading [32, 31], wireless communication networks [12, 11, 13], and learning behaviors in social dilemmas [33, 28, 34]. MARL, however, becomes intractable due to the complex interactions among agents as the number of agents increases. A recent tractable approach is a mean-field approach by considering MARL in the regime with a large number of homogeneous agents under weak interactions [20]. According to the number of agents and learning goals, there are three subtle types of mean-field theories for MARL. The first one is called mean-field MARL (MF-MARL), which refers to the empirical average of the states or actions of a finite population. For example, [52] proposes to approximate interactions within the population of agents by averaging the actions of the overall population or neighboring agents.

Deep Reinforcement Learning (DRL) has made tremendous advances in both simulated and real-world robot control tasks in recent years. Nevertheless, applying DRL to novel robot control tasks is still challenging, especially when researchers have to design the action and observation space and the reward function. In this paper, we investigate partial observability as a potential failure source of applying DRL to robot control tasks, which can occur when researchers are not confident whether the observation space fully represents the underlying state. We compare the performance of three common DRL algorithms, TD3, SAC and PPO under various partial observability conditions. We find that TD3 and SAC become easily stuck in local optima and underperform PPO. We propose multi-step versions of the vanilla TD3 and SAC to improve robustness to partial observability based on one-step bootstrapping.

Deep reinforcement learning is a technique for solving problems in a variety of environments, ranging from Atari video games to stock trading. This method leverages deep neural network models to make decisions based on observations of a given environment with the goal of maximizing a reward function that can incorporate cost and rewards for reaching goals. With the aim of pathfinding, reward conditions can include reaching a specified target area along with costs for movement. In this work, multiple Deep Q-Network (DQN) agents are trained to operate in a partially observable environment with the goal of reaching a target zone in minimal travel time. The agent operates based on a visual representation of its surroundings, and thus has a restricted capability to observe the environment. A comparison between DQN, DQN-GRU, and DQN-LSTM is performed to examine each models capabilities with two different types of input. Through this evaluation, it is been shown that with equivalent training and analogous model architectures, a DQN model is able to outperform its recurrent counterparts.

The recent advancement of Deep Reinforcement Learning (DRL) contributed to robotics by allowing automatic controller design. The automatic controller design is a crucial approach for designing swarm robotic systems, which require more complex controllers than a single robot system to lead a desired collective behaviour. Although the DRL-based controller design method showed its effectiveness, the reliance on the central training server is a critical problem in real-world environments where robot-server communication is unstable or limited. We propose a novel Federated Learning (FL) based DRL training strategy (FLDDPG) for use in swarm robotic applications. Through the comparison with baseline strategies under a limited communication bandwidth scenario, it is shown that the FLDDPG method resulted in higher robustness and generalisation ability into a different environment and real robots, while the baseline strategies suffer from the limitation of communication bandwidth. This result suggests that the proposed method can benefit swarm robotic systems operating in environments with limited communication bandwidth, e.g., in high-radiation, underwater, or subterranean environments.

As safety violations can lead to severe consequences in real-world robotic applications, the increasing deployment of Reinforcement Learning (RL) in robotic domains has propelled the study of safe exploration for reinforcement learning (safe RL). In this work, we propose a risk preventive training method for safe RL, which learns a statistical contrastive classifier to predict the probability of a state-action pair leading to unsafe states. Based on the predicted risk probabilities, we can collect risk preventive trajectories and reshape the reward function with risk penalties to induce safe RL policies. We conduct experiments in robotic simulation environments. The results show the proposed approach has comparable performance with the state-of-the-art model-based methods and outperforms conventional model-free safe RL approaches.

Off-policy algorithms, in which a behavior policy differs from the target policy and is used to gain experience for learning, have proven to be of great practical value in reinforcement learning. However, even for simple convex problems such as linear value function approximation, these algorithms are not guaranteed to be stable. To address this, alternative algorithms that are provably convergent in such cases have been introduced, the most well known being gradient descent temporal difference (GTD) learning. This algorithm and others like it, however, tend to converge much more slowly than conventional temporal difference learning. In this paper we propose gradient descent temporal difference-difference (Gradient-DD) learning in order to improve GTD2, a GTD algorithm, by introducing second-order differences in successive parameter updates. We investigate this algorithm in the framework of linear value function approximation, theoretically proving its convergence by applying the theory of stochastic approximation. %analytically showing its improvement over GTD2. Studying the model empirically on the random walk task, the Boyan-chain task, and the Baird's off-policy counterexample, we find substantial improvement over GTD2 and, in several cases, better performance even than conventional TD learning.

Vision-Language Navigation requires the agent to follow natural language instructions to reach a specific target. The large discrepancy between seen and unseen environments makes it challenging for the agent to generalize well. Previous studies propose data augmentation methods to mitigate the data bias explicitly or implicitly and provide improvements in generalization. However, they try to memorize augmented trajectories and ignore the distribution shifts under unseen environments at test time. In this paper, we propose an Unseen Discrepancy Anticipating Vision and Language Navigation (DAVIS) that learns to generalize to unseen environments via encouraging test-time visual consistency. Specifically, we devise: 1) a semi-supervised framework DAVIS that leverages visual consistency signals across similar semantic observations. 2) a two-stage learning procedure that encourages adaptation to test-time distribution. The framework enhances the basic mixture of imitation and reinforcement learning with Momentum Contrast to encourage stable decision-making on similar observations under a joint training stage and a test-time adaptation stage. Extensive experiments show that DAVIS achieves model-agnostic improvement over previous state-of-the-art VLN baselines on R2R and RxR benchmarks. Our source code and data are in supplemental materials.

Joint caching and transmission optimization problem is challenging due to the deep coupling between decisions. This paper proposes an iterative distributed multi-agent learning approach to jointly optimize caching and transmission. The goal of this approach is to minimize the total transmission delay of all users. In this iterative approach, each iteration includes caching optimization and transmission optimization. A multi-agent reinforcement learning (MARL)-based caching network is developed to cache popular tasks, such as answering which files to evict from the cache and which files to storage. Based on the cached files of the caching network, the transmission network transmits cached files for users by single transmission (ST) or joint transmission (JT) with multi-agent Bayesian learning automaton (MABLA) method. And then users access the edge servers with the minimum transmission delay. The experimental results demonstrate the performance of the proposed multi-agent learning approach.

This paper addresses a safe planning and control problem for mobile robots operating in communication- and sensor-limited dynamic environments. In this case the robots cannot sense the objects around them and must instead rely on intermittent, external information about the environment, as e.g., in underwater applications. The challenge in this case is that the robots must plan using only this stale data, while accounting for any noise in the data or uncertainty in the environment. To address this challenge we propose a compositional technique which leverages neural networks to quickly plan and control a robot through crowded and dynamic environments using only intermittent information. Specifically, our tool uses reachability analysis and potential fields to train a neural network that is capable of generating safe control actions. We demonstrate our technique both in simulation with an underwater vehicle crossing a crowded shipping channel and with real experiments with ground vehicles in communication- and sensor-limited environments.

His areas of expertise are deep reinforcement learning and information systems design.