In multi-task reinforcement learning, agents train on a fixed set of tasks and have to generalise to new ones. Recent work has shown that increased exploration improves this generalisation, but it remains unclear why exactly that is. In this paper, we introduce the concept of reachability in multi-task reinforcement learning and show that an initial exploration phase increases the number of reachable tasks the agent is trained on. This, and not the increased exploration, is responsible for the improved generalisation, even to unreachable tasks. Inspired by this, we propose a novel method Explore-Go that implements such an exploration phase at the beginning of each episode. Explore-Go only modifies the way experience is collected and can be used with most existing on-policy or off-policy reinforcement learning algorithms. We demonstrate the effectiveness of our method when combined with some popular algorithms and show an increase in generalisation performance across several environments.
Recent studies reveal that well-performing reinforcement learning (RL) agents in training often lack resilience against adversarial perturbations during deployment. This highlights the importance of building a robust agent before deploying it in the real world. Most prior works focus on developing robust training-based procedures to tackle this problem, including enhancing the robustness of the deep neural network component itself or adversarially training the agent on strong attacks. In this work, we instead study an input transformation-based defense for RL. Specifically, we propose using a variant of vector quantization (VQ) as a transformation for input observations, which is then used to reduce the space of adversarial attacks during testing, resulting in the transformed observations being less affected by attacks. Our method is computationally efficient and seamlessly integrates with adversarial training, further enhancing the robustness of RL agents against adversarial attacks. Through extensive experiments in multiple environments, we demonstrate that using VQ as the input transformation effectively defends against adversarial attacks on the agent's observations.
Deep Reinforcement Learning (DRL) enables robots to learn complex behaviors through interaction with the environment. However, due to the unrestricted nature of the learning algorithms, the resulting solutions are often brittle and appear unnatural. This is especially true for learning direct joint-level torque control, as inductive biases are difficult to integrate into the learning process. We propose an inductive bias for learning locomotion that is especially useful for torque control: latent actions learned from a small dataset of expert demonstrations. This prior allows the policy to directly leverage knowledge contained in the expert's actions and facilitates more efficient exploration. We observe that the agent is not restricted to the reward levels of the demonstration, and performance in transfer tasks is improved significantly. Latent action priors combined with style rewards for imitation lead to a closer replication of the expert's behavior. Videos and code are available at https://sites.google.com/view/latent-action-priors.
Reinforcement learning from human feedback (RLHF) has demonstrated effectiveness in aligning large language models (LLMs) with human preferences. However, token-level RLHF suffers from the credit assignment problem over long sequences, where delayed rewards make it challenging for the model to discern which actions contributed to successful outcomes. This hinders learning efficiency and slows convergence. In this paper, we propose MA-RLHF, a simple yet effective RLHF framework that incorporates macro actions -- sequences of tokens or higher-level language constructs -- into the learning process. By operating at this higher level of abstraction, our approach reduces the temporal distance between actions and rewards, facilitating faster and more accurate credit assignment. This results in more stable policy gradient estimates and enhances learning efficiency within each episode, all without increasing computational complexity during training or inference. We validate our approach through extensive experiments across various model sizes and tasks, including text summarization, dialogue generation, question answering, and program synthesis. Our method achieves substantial performance improvements over standard RLHF, with performance gains of up to 30% in text summarization and code generation, 18% in dialogue, and 8% in question answering tasks. Notably, our approach reaches parity with vanilla RLHF 1.7x to 2x faster in terms of training time and continues to outperform it with further training. We will make our code and data publicly available at https://github.com/ernie-research/MA-RLHF .
While Deep Reinforcement Learning (DRL) has achieved remarkable success across various domains, it remains vulnerable to occasional catastrophic failures without additional safeguards. An effective solution to prevent these failures is to use a shield that validates and adjusts the agent's actions to ensure compliance with a provided set of safety specifications. For real-world robotic domains, it is essential to define safety specifications over continuous state and action spaces to accurately account for system dynamics and compute new actions that minimally deviate from the agent's original decision. In this paper, we present the first shielding approach specifically designed to ensure the satisfaction of safety requirements in continuous state and action spaces, making it suitable for practical robotic applications. Our method builds upon realizability, an essential property that confirms the shield will always be able to generate a safe action for any state in the environment. We formally prove that realizability can be verified for stateful shields, enabling the incorporation of non-Markovian safety requirements, such as loop avoidance. Finally, we demonstrate the effectiveness of our approach in ensuring safety without compromising the policy's success rate by applying it to a navigation problem and a multi-agent particle environment.
Offline reinforcement learning (RL) has garnered significant attention for its ability to learn effective policies from pre-collected datasets without the need for further environmental interactions. While promising results have been demonstrated in single-agent settings, offline multi-agent reinforcement learning (MARL) presents additional challenges due to the large joint state-action space and the complexity of multi-agent behaviors. A key issue in offline RL is the distributional shift, which arises when the target policy being optimized deviates from the behavior policy that generated the data. This problem is exacerbated in MARL due to the interdependence between agents' local policies and the expansive joint state-action space. Prior approaches have primarily addressed this challenge by incorporating regularization in the space of either Q-functions or policies. In this work, we introduce a regularizer in the space of stationary distributions to better handle distributional shift. Our algorithm, ComaDICE, offers a principled framework for offline cooperative MARL by incorporating stationary distribution regularization for the global learning policy, complemented by a carefully structured multi-agent value decomposition strategy to facilitate multi-agent training. Through extensive experiments on the multi-agent MuJoCo and StarCraft II benchmarks, we demonstrate that ComaDICE achieves superior performance compared to state-of-the-art offline MARL methods across nearly all tasks.
Policy evaluation estimates the performance of a policy by (1) collecting data from the environment and (2) processing raw data into a meaningful estimate. Due to the sequential nature of reinforcement learning, any improper data-collecting policy or data-processing method substantially deteriorates the variance of evaluation results over long time steps. Thus, policy evaluation often suffers from large variance and requires massive data to achieve the desired accuracy. In this work, we design an optimal combination of data-collecting policy and data-processing baseline. Theoretically, we prove our doubly optimal policy evaluation method is unbiased and guaranteed to have lower variance than previously best-performing methods. Empirically, compared with previous works, we show our method reduces variance substantially and achieves superior empirical performance.
Aligning large language models (LLMs) with human preferences is critical to recent advances in generative artificial intelligence. Reinforcement learning from human feedback (RLHF) is widely applied to achieve this objective. A key step in RLHF is to learn the reward function from human feedback. However, human feedback is costly and time-consuming, making it essential to collect high-quality conversation data for human teachers to label. Additionally, different human teachers have different levels of expertise. It is thus critical to query the most appropriate teacher for their opinions. In this paper, we use offline reinforcement learning (RL) to formulate the alignment problem. Motivated by the idea of $D$-optimal design, we first propose a dual active reward learning algorithm for the simultaneous selection of conversations and teachers. Next, we apply pessimistic RL to solve the alignment problem, based on the learned reward estimator. Theoretically, we show that the reward estimator obtained through our proposed adaptive selection strategy achieves minimal generalized variance asymptotically, and prove that the sub-optimality of our pessimistic policy scales as $O(1/\sqrt{T})$ with a given sample budget $T$. Through simulations and experiments on LLMs, we demonstrate the effectiveness of our algorithm and its superiority over state-of-the-arts.
Dexterous hands exhibit significant potential for complex real-world grasping tasks. While recent studies have primarily focused on learning policies for specific robotic hands, the development of a universal policy that controls diverse dexterous hands remains largely unexplored. In this work, we study the learning of cross-embodiment dexterous grasping policies using reinforcement learning (RL). Inspired by the capability of human hands to control various dexterous hands through teleoperation, we propose a universal action space based on the human hand's eigengrasps. The policy outputs eigengrasp actions that are then converted into specific joint actions for each robot hand through a retargeting mapping. We simplify the robot hand's proprioception to include only the positions of fingertips and the palm, offering a unified observation space across different robot hands. Our approach demonstrates an 80% success rate in grasping objects from the YCB dataset across four distinct embodiments using a single vision-based policy. Additionally, our policy exhibits zero-shot generalization to two previously unseen embodiments and significant improvement in efficient finetuning. For further details and videos, visit our project page https://sites.google.com/view/crossdex.
Large language models (LLMs) deployed as agents solve user-specified tasks over multiple steps while keeping the required manual engagement to a minimum. Crucially, such LLMs need to ground their generations in any feedback obtained to reliably achieve desired outcomes. We propose an end-to-end reinforcement learning method for teaching models to leverage execution feedback in the realm of code synthesis, where state-of-the-art LLMs struggle to improve code iteratively compared to independent sampling. We benchmark on competitive programming tasks, where we achieve new start-of-the art results with both small (8B parameters) and large (70B) models while reducing the amount of samples required by an order of magnitude. Our analysis of inference-time behavior demonstrates that our method produces LLMs that effectively leverage automatic feedback over multiple steps.
One of the fundamental challenges in reinforcement learning (RL) is to take a complex task and be able to decompose it to subtasks that are simpler for the RL agent to learn. In this paper, we report on our work that would identify subtasks by using some given positive and negative trajectories for solving the complex task. We assume that the states are represented by first-order predicate logic using which we devise a novel algorithm to identify the subtasks. Then we employ a Large Language Model (LLM) to generate first-order logic rule templates for achieving each subtask. Such rules were then further fined tuned to a rule-based policy via an Inductive Logic Programming (ILP)-based RL agent. Through experiments, we verify the accuracy of our algorithm in detecting subtasks which successfully detect all of the subtasks correctly. We also investigated the quality of the common-sense rules produced by the language model to achieve the subtasks. Our experiments show that our LLM-guided rule template generation can produce rules that are necessary for solving a subtask, which leads to solving complex tasks with fewer assumptions about predefined first-order logic predicates of the environment.
The deep reinforcement learning (DRL) based Volt-VAR optimization (VVO) methods have been widely studied for active distribution networks (ADNs). However, most of them lack safety guarantees in terms of power injection uncertainties due to the increase in distributed energy resources (DERs) and load demand, such as electric vehicles. This article proposes a robust deep reinforcement learning (RDRL) framework for VVO via a robust deep deterministic policy gradient (DDPG) algorithm. This algorithm can effectively manage hybrid action spaces, considering control devices like capacitors, voltage regulators, and smart inverters. Additionally, it is designed to handle uncertainties by quantifying uncertainty sets with conformal prediction and modeling uncertainties as adversarial attacks to guarantee safe exploration across action spaces. Numerical results on three IEEE test cases demonstrate the sample efficiency and safety of the proposed robust DDPG against uncertainties compared to the benchmark algorithms.
Greenhouse climate control is concerned with maximizing performance in terms of crop yield and resource efficiency. One promising approach is model predictive control (MPC), which leverages a model of the system to optimize the control inputs, while enforcing physical constraints. However, prediction models for greenhouse systems are inherently inaccurate due to the complexity of the real system and the uncertainty in predicted weather profiles. For model-based control approaches such as MPC, this can degrade performance and lead to constraint violations. Existing approaches address uncertainty in the prediction model with robust or stochastic MPC methodology; however, these necessarily reduce crop yield due to conservatism and often bear higher computational loads. In contrast, learning-based control approaches, such as reinforcement learning (RL), can handle uncertainty naturally by leveraging data to improve performance. This work proposes an MPC-based RL control framework to optimize the climate control performance in the presence of prediction uncertainty. The approach employs a parametrized MPC scheme that learns directly from data, in an online fashion, the parametrization of the constraints, prediction model, and optimization cost that minimizes constraint violations and maximizes climate control performance. Simulations show that the approach can learn an MPC controller that significantly outperforms the current state-of-the-art in terms of constraint violations and efficient crop growth.
The next-generation wireless network, 6G and beyond, envisions to integrate communication and sensing to overcome interference, improve spectrum efficiency, and reduce hardware and power consumption. Massive Multiple-Input Multiple Output (mMIMO)-based Joint Communication and Sensing (JCAS) systems realize this integration for 6G applications such as autonomous driving, as it requires accurate environmental sensing and time-critical communication with neighboring vehicles. Reinforcement Learning (RL) is used for mMIMO antenna beamforming in the existing literature. However, the huge search space for actions associated with antenna beamforming causes the learning process for the RL agent to be inefficient due to high beam training overhead. The learning process does not consider the causal relationship between action space and the reward, and gives all actions equal importance. In this work, we explore a causally-aware RL agent which can intervene and discover causal relationships for mMIMO-based JCAS environments, during the training phase. We use a state dependent action dimension selection strategy to realize causal discovery for RL-based JCAS. Evaluation of the causally-aware RL framework in different JCAS scenarios shows the benefit of our proposed framework over baseline methods in terms of the beamforming gain.
Real-time dynamic scheduling is a crucial but notoriously challenging task in modern manufacturing processes due to its high decision complexity. Recently, reinforcement learning (RL) has been gaining attention as an impactful technique to handle this challenge. However, classical RL methods typically rely on human-made dispatching rules, which are not suitable for large-scale factory-wide scheduling. To bridge this gap, this paper applies a leader-follower multi-agent RL (MARL) concept to obtain desired coordination after decomposing the scheduling problem into a set of sub-problems that are handled by each individual agent for scalability. We further strengthen the procedure by proposing a rule-based conversion algorithm to prevent catastrophic loss of production capacity due to an agent's error. Our experimental results demonstrate that the proposed model outperforms the state-of-the-art deep RL-based scheduling models in various aspects. Additionally, the proposed model provides the most robust scheduling performance to demand changes. Overall, the proposed MARL-based scheduling model presents a promising solution to the real-time scheduling problem, with potential applications in various manufacturing industries.
In this publication, we assess the ability of a novel Reinforcement Learning-based solution to the problem of Neural Architecture Search, where a Reinforcement Learning (RL) agent learns to search for good architectures, rather than to return a single optimal architecture. We consider both the NAS-Bench-101 and NAS- Bench-301 settings, and compare against various known strong baselines, such as local search and random search. We conclude that our Reinforcement Learning agent displays strong scalability with regards to the size of the search space, but limited robustness to hyperparameter changes.
Intrinsic motivation, inspired by the psychology of developmental learning in infants, stimulates exploration in agents without relying solely on sparse external rewards. Existing methods in reinforcement learning like Random Network Distillation (RND) face significant limitations, including (1) relying on raw visual inputs, leading to a lack of meaningful representations, (2) the inability to build a robust latent space, (3) poor target network initialization and (4) rapid degradation of intrinsic rewards. In this paper, we introduce Pre-trained Network Distillation (PreND), a novel approach to enhance intrinsic motivation in reinforcement learning (RL) by improving upon the widely used prediction-based method, RND. PreND addresses these challenges by incorporating pre-trained representation models into both the target and predictor networks, resulting in more meaningful and stable intrinsic rewards, while enhancing the representation learned by the model. We also tried simple but effective variants of the predictor network optimization by controlling the learning rate. Through experiments on the Atari domain, we demonstrate that PreND significantly outperforms RND, offering a more robust intrinsic motivation signal that leads to better exploration, improving overall performance and sample efficiency. This research highlights the importance of target and predictor networks representation in prediction-based intrinsic motivation, setting a new direction for improving RL agents' learning efficiency in sparse reward environments.
Learning agents with reinforcement learning is difficult when dealing with long trajectories that involve a large number of states. To address these learning problems effectively, the number of states can be reduced by abstract representations that cluster states. In principle, deep reinforcement learning can find abstract states, but end-to-end learning is unstable. We propose contrastive abstraction learning to find abstract states, where we assume that successive states in a trajectory belong to the same abstract state. Such abstract states may be basic locations, achieved subgoals, inventory, or health conditions. Contrastive abstraction learning first constructs clusters of state representations by contrastive learning and then applies modern Hopfield networks to determine the abstract states. The first phase of contrastive abstraction learning is self-supervised learning, where contrastive learning forces states with sequential proximity to have similar representations. The second phase uses modern Hopfield networks to map similar state representations to the same fixed point, i.e.\ to an abstract state. The level of abstraction can be adjusted by determining the number of fixed points of the modern Hopfield network. Furthermore, \textit{contrastive abstraction learning} does not require rewards and facilitates efficient reinforcement learning for a wide range of downstream tasks. Our experiments demonstrate the effectiveness of contrastive abstraction learning for reinforcement learning.
As multi-agent reinforcement learning (MARL) progresses towards solving larger and more complex problems, it becomes increasingly important that algorithms exhibit the key properties of (1) strong performance, (2) memory efficiency, and (3) scalability. In this work, we introduce Sable, a performant, memory-efficient, and scalable sequence modeling approach to MARL. Sable works by adapting the retention mechanism in Retentive Networks (Sun et al., 2023) to achieve computationally efficient processing of multi-agent observations with long context memory for temporal reasoning. Through extensive evaluations across six diverse environments, we demonstrate how Sable is able to significantly outperform existing state-of-the-art methods in a large number of diverse tasks (34 out of 45 tested). Furthermore, Sable maintains performance as we scale the number of agents, handling environments with more than a thousand agents while exhibiting a linear increase in memory usage. Finally, we conduct ablation studies to isolate the source of Sable's performance gains and confirm its efficient computational memory usage.
This paper presents Grammar Reinforcement Learning (GRL), a reinforcement learning algorithm that uses Monte Carlo Tree Search (MCTS) and a transformer architecture that models a Pushdown Automaton (PDA) within a context-free grammar (CFG) framework. Taking as use case the problem of efficiently counting paths and cycles in graphs, a key challenge in network analysis, computer science, biology, and social sciences, GRL discovers new matrix-based formulas for path/cycle counting that improve computational efficiency by factors of two to six w.r.t state-of-the-art approaches. Our contributions include: (i) a framework for generating gramformers that operate within a CFG, (ii) the development of GRL for optimizing formulas within grammatical structures, and (iii) the discovery of novel formulas for graph substructure counting, leading to significant computational improvements.