Incorporating safety is an essential prerequisite for broadening the practical applications of reinforcement learning in real-world scenarios. To tackle this challenge, Constrained Markov Decision Processes (CMDPs) are leveraged, which introduce a distinct cost function representing safety violations. In CMDPs' settings, Lagrangian relaxation technique has been employed in previous algorithms to convert constrained optimization problems into unconstrained dual problems. However, these algorithms may inaccurately predict unsafe behavior, resulting in instability while learning the Lagrange multiplier. This study introduces a novel safe reinforcement learning algorithm, Safety Critic Policy Optimization (SCPO). In this study, we define the safety critic, a mechanism that nullifies rewards obtained through violating safety constraints. Furthermore, our theoretical analysis indicates that the proposed algorithm can automatically balance the trade-off between adhering to safety constraints and maximizing rewards. The effectiveness of the SCPO algorithm is empirically validated by benchmarking it against strong baselines.

We present a novel multi-agent RL approach, Selective Multi-Agent Prioritized Experience Relay, in which agents share with other agents a limited number of transitions they observe during training. The intuition behind this is that even a small number of relevant experiences from other agents could help each agent learn. Unlike many other multi-agent RL algorithms, this approach allows for largely decentralized training, requiring only a limited communication channel between agents. We show that our approach outperforms baseline no-sharing decentralized training and state-of-the art multi-agent RL algorithms. Further, sharing only a small number of highly relevant experiences outperforms sharing all experiences between agents, and the performance uplift from selective experience sharing is robust across a range of hyperparameters and DQN variants. A reference implementation of our algorithm is available at https://github.com/mgerstgrasser/super.

In the field of explainable Artificial Intelligence (XAI), sequential counterfactual (SCF) examples are often used to alter the decision of a trained classifier by implementing a sequence of modifications to the input instance. Although certain test-time algorithms aim to optimize for each new instance individually, recently Reinforcement Learning (RL) methods have been proposed that seek to learn policies for discovering SCFs, thereby enhancing scalability. As is typical in RL, the formulation of the RL problem, including the specification of state space, actions, and rewards, can often be ambiguous. In this work, we identify shortcomings in existing methods that can result in policies with undesired properties, such as a bias towards specific actions. We propose to use the output probabilities of the classifier to create a more informative reward, to mitigate this effect.

Exploration poses a fundamental challenge in Reinforcement Learning (RL) with sparse rewards, limiting an agent's ability to learn optimal decision-making due to a lack of informative feedback signals. Self-Imitation Learning (self-IL) has emerged as a promising approach for exploration, leveraging a replay buffer to store and reproduce successful behaviors. However, traditional self-IL methods, which rely on high-return transitions and assume singleton environments, face challenges in generalization, especially in procedurally-generated (PCG) environments. Therefore, new self-IL methods have been proposed to rank which experiences to persist, but they replay transitions uniformly regardless of their significance, and do not address the diversity of the stored demonstrations. In this work, we propose tailored self-IL sampling strategies by prioritizing transitions in different ways and extending prioritization techniques to PCG environments. We also address diversity loss through modifications to counteract the impact of generalization requirements and bias introduced by prioritization techniques. Our experimental analysis, conducted over three PCG sparse reward environments, including MiniGrid and ProcGen, highlights the benefits of our proposed modifications, achieving a new state-of-the-art performance in the MiniGrid-MultiRoom-N12-S10 environment.

Centralized training is widely utilized in the field of multi-agent reinforcement learning (MARL) to assure the stability of training process. Once a joint policy is obtained, it is critical to design a value function factorization method to extract optimal decentralized policies for the agents, which needs to satisfy the individual-global-max (IGM) principle. While imposing additional limitations on the IGM function class can help to meet the requirement, it comes at the cost of restricting its application to more complex multi-agent environments. In this paper, we propose QFree, a universal value function factorization method for MARL. We start by developing mathematical equivalent conditions of the IGM principle based on the advantage function, which ensures that the principle holds without any compromise, removing the conservatism of conventional methods. We then establish a more expressive mixing network architecture that can fulfill the equivalent factorization. In particular, the novel loss function is developed by considering the equivalent conditions as regularization term during policy evaluation in the MARL algorithm. Finally, the effectiveness of the proposed method is verified in a nonmonotonic matrix game scenario. Moreover, we show that QFree achieves the state-of-the-art performance in a general-purpose complex MARL benchmark environment, Starcraft Multi-Agent Challenge (SMAC).

Recent years have witnessed many successful trials in the robot learning field. For contact-rich robotic tasks, it is challenging to learn coordinated motor skills by reinforcement learning. Imitation learning solves this problem by using a mimic reward to encourage the robot to track a given reference trajectory. However, imitation learning is not so efficient and may constrain the learned motion. In this paper, we propose instruction learning, which is inspired by the human learning process and is highly efficient, flexible, and versatile for robot motion learning. Instead of using a reference signal in the reward, instruction learning applies a reference signal directly as a feedforward action, and it is combined with a feedback action learned by reinforcement learning to control the robot. Besides, we propose the action bounding technique and remove the mimic reward, which is shown to be crucial for efficient and flexible learning. We compare the performance of instruction learning with imitation learning, indicating that instruction learning can greatly speed up the training process and guarantee learning the desired motion correctly. The effectiveness of instruction learning is validated through a bunch of motion learning examples for a biped robot and a quadruped robot, where skills can be learned typically within several million steps. Besides, we also conduct sim-to-real transfer and online learning experiments on a real quadruped robot. Instruction learning has shown great merits and potential, making it a promising alternative for imitation learning.

Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedback-trained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality. Our code is available at https://github.com/google-research/google-research/tree/master/dpok.

Energy market rules should incentivize market participants to behave in a market and grid conform way. However, they can also provide incentives for undesired and unexpected strategies if the market design is flawed. Multi-agent Reinforcement learning (MARL) is a promising new approach to predicting the expected profit-maximizing behavior of energy market participants in simulation. However, reinforcement learning requires many interactions with the system to converge, and the power system environment often consists of extensive computations, e.g., optimal power flow (OPF) calculation for market clearing. To tackle this complexity, we provide a model of the energy market to a basic MARL algorithm in the form of a learned OPF approximation and explicit market rules. The learned OPF surrogate model makes an explicit solving of the OPF completely unnecessary. Our experiments demonstrate that the model additionally reduces training time by about one order of magnitude but at the cost of a slightly worse performance. Potential applications of our method are market design, more realistic modeling of market participants, and analysis of manipulative behavior.

Reinforcement Learning (RL) is a popular machine learning paradigm where intelligent agents interact with the environment to fulfill a long-term goal. Driven by the resurgence of deep learning, Deep RL (DRL) has witnessed great success over a wide spectrum of complex control tasks. Despite the encouraging results achieved, the deep neural network-based backbone is widely deemed as a black box that impedes practitioners to trust and employ trained agents in realistic scenarios where high security and reliability are essential. To alleviate this issue, a large volume of literature devoted to shedding light on the inner workings of the intelligent agents has been proposed, by constructing intrinsic interpretability or post-hoc explainability. In this survey, we provide a comprehensive review of existing works on eXplainable RL (XRL) and introduce a new taxonomy where prior works are clearly categorized into model-explaining, reward-explaining, state-explaining, and task-explaining methods. We also review and highlight RL methods that conversely leverage human knowledge to promote learning efficiency and performance of agents while this kind of method is often ignored in XRL field. Some challenges and opportunities in XRL are discussed. This survey intends to provide a high-level summarization of XRL and to motivate future research on more effective XRL solutions. Corresponding open source codes are collected and categorized at https://github.com/Plankson/awesome-explainable-reinforcement-learning.

We provide a new perspective to understand why reinforcement learning (RL) struggles with robustness and generalization. We show, by examples, that local optimal policies may contain unstable control for some dynamic parameters and overfitting to such instabilities can deteriorate robustness and generalization. Contraction analysis of neural control reveals that there exists boundaries between stable and unstable control with respect to the input gradients of control networks. Ignoring those stability boundaries, learning agents may label the actions that cause instabilities for some dynamic parameters as high value actions if those actions can improve the expected return. The small fraction of such instabilities may not cause attention in the empirical studies, a hidden risk for real-world applications. Those instabilities can manifest themselves via overfitting, leading to failures in robustness and generalization. We propose stability constraints and terminal constraints to solve this issue, demonstrated with a proximal policy optimization example.