For autonomous agents to successfully integrate into human-centered environments, agents should be able to learn from and adapt to humans in their native settings. Preference-based reinforcement learning (PbRL) is a promising approach that learns reward functions from human preferences. This enables RL agents to adapt their behavior based on human desires. However, humans live in a world full of diverse information, most of which is not relevant to completing a particular task. It becomes essential that agents learn to focus on the subset of task-relevant environment features. Unfortunately, prior work has largely ignored this aspect; primarily focusing on improving PbRL algorithms in standard RL environments that are carefully constructed to contain only task-relevant features. This can result in algorithms that may not effectively transfer to a more noisy real-world setting. To that end, this work proposes R2N (Robust-to-Noise), the first PbRL algorithm that leverages principles of dynamic sparse training to learn robust reward models that can focus on task-relevant features. We study the effectiveness of R2N in the Extremely Noisy Environment setting, an RL problem setting where up to 95% of the state features are irrelevant distractions. In experiments with a simulated teacher, we demonstrate that R2N can adapt the sparse connectivity of its neural networks to focus on task-relevant features, enabling R2N to significantly outperform several state-of-the-art PbRL algorithms in multiple locomotion and control environments.
Markov games (MGs) and multi-agent reinforcement learning (MARL) are studied to model decision making in multi-agent systems. Traditionally, the objective in MG and MARL has been risk-neutral, i.e., agents are assumed to optimize a performance metric such as expected return, without taking into account subjective or cognitive preferences of themselves or of other agents. However, ignoring such preferences leads to inaccurate models of decision making in many real-world scenarios in finance, operations research, and behavioral economics. Therefore, when these preferences are present, it is necessary to incorporate a suitable measure of risk into the optimization objective of agents, which opens the door to risk-sensitive MG and MARL. In this paper, we systemically review the literature on risk sensitivity in MG and MARL that has been growing in recent years alongside other areas of reinforcement learning and game theory. We define and mathematically describe different risk measures used in MG and MARL and individually for each measure, discuss articles that incorporate it. Finally, we identify recent trends in theoretical and applied works in the field and discuss possible directions of future research.
Developing interactive systems that utilize natural language instructions to solve complex robotic control tasks has long been a goal of the robotics community. While Large Language Models (LLMs) excel at logical reasoning, in-context learning, and code generation, translating high-level instructions into low-level robotic actions still remains challenging. Furthermore, solving such tasks often requires acquiring policies to execute diverse subtasks and integrating them to achieve the final objective. Hierarchical Reinforcement Learning (HRL) offers a promising solution for solving such tasks by enabling temporal abstraction and improved exploration. However, HRL suffers from non-stationarity caused by the changing lower-level behaviour, which hinders effective policy learning. We propose LGR2, a novel HRL framework that mitigates non-stationarity in HRL by using language-guided higher-level rewards that remain unaffected by the changing lower-level policy behaviour. To analyze the efficacy of our approach, we perform empirical analysis to demonstrate that LGR2 effectively mitigates non-stationarity in HRL and attains success rates exceeding 70% in challenging, sparsely-rewarded robotic navigation and manipulation environments, where other baselines typically fail to show significant progress. Finally, we perform real-world robotic experiments on complex tasks and demonstrate that LGR2 consistently outperforms the baselines.
In this research, we introduce a deep reinforcement learning-based control approach to address the intricate challenge of the robotic pre-grasping phase under microgravity conditions. Leveraging reinforcement learning eliminates the necessity for manual feature design, therefore simplifying the problem and empowering the robot to learn pre-grasping policies through trial and error. Our methodology incorporates an off-policy reinforcement learning framework, employing the soft actor-critic technique to enable the gripper to proficiently approach a free-floating moving object, ensuring optimal pre-grasp success. For effective learning of the pre-grasping approach task, we developed a reward function that offers the agent clear and insightful feedback. Our case study examines a pre-grasping task where a Robotiq 3F gripper is required to navigate towards a free-floating moving target, pursue it, and subsequently position itself at the desired pre-grasp location. We assessed our approach through a series of experiments in both simulated and real-world environments. The source code, along with recordings of real-world robot grasping, is available at Fanuc_Robotiq_Grasp.
In this paper, we introduce Multi-Objective Deep Centralized Multi-Agent Actor-Critic (MO- DCMAC), a multi-objective reinforcement learning (MORL) method for infrastructural maintenance optimization, an area traditionally dominated by single-objective reinforcement learning (RL) approaches. Previous single-objective RL methods combine multiple objectives, such as probability of collapse and cost, into a singular reward signal through reward-shaping. In contrast, MO-DCMAC can optimize a policy for multiple objectives directly, even when the utility function is non-linear. We evaluated MO-DCMAC using two utility functions, which use probability of collapse and cost as input. The first utility function is the Threshold utility, in which MO-DCMAC should minimize cost so that the probability of collapse is never above the threshold. The second is based on the Failure Mode, Effects, and Criticality Analysis (FMECA) methodology used by asset managers to asses maintenance plans. We evaluated MO-DCMAC, with both utility functions, in multiple maintenance environments, including ones based on a case study of the historical quay walls of Amsterdam. The performance of MO-DCMAC was compared against multiple rule-based policies based on heuristics currently used for constructing maintenance plans. Our results demonstrate that MO-DCMAC outperforms traditional rule-based policies across various environments and utility functions.
The expected regret of any reinforcement learning algorithm is lower bounded by $\Omega\left(\sqrt{DXAT}\right)$ for undiscounted returns, where $D$ is the diameter of the Markov decision process, $X$ the size of the state space, $A$ the size of the action space and $T$ the number of time steps. However, this lower bound is general. A smaller regret can be obtained by taking into account some specific knowledge of the problem structure. In this article, we consider an admission control problem to an $M/M/c/S$ queue with $m$ job classes and class-dependent rewards and holding costs. Queuing systems often have a diameter that is exponential in the buffer size $S$, making the previous lower bound prohibitive for any practical use. We propose an algorithm inspired by UCRL2, and use the structure of the problem to upper bound the expected total regret by $O(S\log T + \sqrt{mT \log T})$ in the finite server case. In the infinite server case, we prove that the dependence of the regret on $S$ disappears.
Computing Jacobians with automatic differentiation is ubiquitous in many scientific domains such as machine learning, computational fluid dynamics, robotics and finance. Even small savings in the number of computations or memory usage in Jacobian computations can already incur massive savings in energy consumption and runtime. While there exist many methods that allow for such savings, they generally trade computational efficiency for approximations of the exact Jacobian. In this paper, we present a novel method to optimize the number of necessary multiplications for Jacobian computation by leveraging deep reinforcement learning (RL) and a concept called cross-country elimination while still computing the exact Jacobian. Cross-country elimination is a framework for automatic differentiation that phrases Jacobian accumulation as ordered elimination of all vertices on the computational graph where every elimination incurs a certain computational cost. We formulate the search for the optimal elimination order that minimizes the number of necessary multiplications as a single player game which is played by an RL agent. We demonstrate that this method achieves up to 33% improvements over state-of-the-art methods on several relevant tasks taken from diverse domains. Furthermore, we show that these theoretical gains translate into actual runtime improvements by providing a cross-country elimination interpreter in JAX that can efficiently execute the obtained elimination orders.
In offline reinforcement learning, in-sample learning methods have been widely used to prevent performance degradation caused by evaluating out-of-distribution actions from the dataset. Extreme Q-learning (XQL) employs a loss function based on the assumption that Bellman error follows a Gumbel distribution, enabling it to model the soft optimal value function in an in-sample manner. It has demonstrated strong performance in both offline and online reinforcement learning settings. However, issues remain, such as the instability caused by the exponential term in the loss function and the risk of the error distribution deviating from the Gumbel distribution. Therefore, we propose Maclaurin Expanded Extreme Q-learning to enhance stability. In this method, applying Maclaurin expansion to the loss function in XQL enhances stability against large errors. This approach involves adjusting the modeled value function between the value function under the behavior policy and the soft optimal value function, thus achieving a trade-off between stability and optimality depending on the order of expansion. It also enables adjustment of the error distribution assumption from a normal distribution to a Gumbel distribution. Our method significantly stabilizes learning in online RL tasks from DM Control, where XQL was previously unstable. Additionally, it improves performance in several offline RL tasks from D4RL.
Meta-Reinforcement Learning (Meta-RL) agents can struggle to operate across tasks with varying environmental features that require different optimal skills (i.e., different modes of behaviour). Using context encoders based on contrastive learning to enhance the generalisability of Meta-RL agents is now widely studied but faces challenges such as the requirement for a large sample size, also referred to as the $\log$-$K$ curse. To improve RL generalisation to different tasks, we first introduce Skill-aware Mutual Information (SaMI), an optimisation objective that aids in distinguishing context embeddings according to skills, thereby equipping RL agents with the ability to identify and execute different skills across tasks. We then propose Skill-aware Noise Contrastive Estimation (SaNCE), a $K$-sample estimator used to optimise the SaMI objective. We provide a framework for equipping an RL agent with SaNCE in practice and conduct experimental validation on modified MuJoCo and Panda-gym benchmarks. We empirically find that RL agents that learn by maximising SaMI achieve substantially improved zero-shot generalisation to unseen tasks. Additionally, the context encoder trained with SaNCE demonstrates greater robustness to a reduction in the number of available samples, thus possessing the potential to overcome the $\log$-$K$ curse.
Deep Reinforcement Learning (DRL) policies are highly susceptible to adversarial noise in observations, which poses significant risks in safety-critical scenarios. For instance, a self-driving car could experience catastrophic consequences if its sensory inputs about traffic signs are manipulated by an adversary. The core challenge in such situations is that the true state of the environment becomes only partially observable due to these adversarial manipulations. Two key strategies have so far been employed in the literature; the first set of methods focuses on increasing the likelihood that nearby states--those close to the true state--share the same robust actions. The second set of approaches maximize the value for the worst possible true state within the range of adversarially perturbed observations. Although these approaches provide strong robustness against attacks, they tend to be either overly conservative or not generalizable. We hypothesize that the shortcomings of these approaches stem from their failure to explicitly account for partial observability. By making decisions that directly consider this partial knowledge of the true state, we believe it is possible to achieve a better balance between robustness and performance, particularly in adversarial settings. To achieve this, we introduce a novel objective called Adversarial Counterfactual Error (ACoE), which is defined on the beliefs about the underlying true state and naturally balances value optimization with robustness against adversarial attacks, and a theoretically-grounded, scalable surrogate objective Cumulative-ACoE (C-ACoE). Our empirical evaluations demonstrate that our method significantly outperforms current state-of-the-art approaches for addressing adversarial RL challenges, offering a promising direction for better DRL under adversarial conditions.
Reinforcement Learning from Human Feedback (RLHF) is popular in large language models (LLMs), whereas traditional Reinforcement Learning (RL) often falls short. Current autonomous driving methods typically utilize either human feedback in machine learning, including RL, or LLMs. Most feedback guides the car agent's learning process (e.g., controlling the car). RLHF is usually applied in the fine-tuning step, requiring direct human "preferences," which are not commonly used in optimizing autonomous driving models. In this research, we innovatively combine RLHF and LLMs to enhance autonomous driving safety. Training a model with human guidance from scratch is inefficient. Our framework starts with a pre-trained autonomous car agent model and implements multiple human-controlled agents, such as cars and pedestrians, to simulate real-life road environments. The autonomous car model is not directly controlled by humans. We integrate both physical and physiological feedback to fine-tune the model, optimizing this process using LLMs. This multi-agent interactive environment ensures safe, realistic interactions before real-world application. Finally, we will validate our model using data gathered from real-life testbeds located in New Jersey and New York City.
Current approaches to model-based offline reinforcement learning often incorporate uncertainty-based reward penalization to address the distributional shift problem. These approaches, commonly known as pessimistic value iteration, use Monte Carlo sampling to estimate the Bellman target to perform temporal difference-based policy evaluation. We find out that the randomness caused by this sampling step significantly delays convergence. We present a theoretical result demonstrating the strong dependency of suboptimality on the number of Monte Carlo samples taken per Bellman target calculation. Our main contribution is a deterministic approximation to the Bellman target that uses progressive moment matching, a method developed originally for deterministic variational inference. The resulting algorithm, which we call Moment Matching Offline Model-Based Policy Optimization (MOMBO), propagates the uncertainty of the next state through a nonlinear Q-network in a deterministic fashion by approximating the distributions of hidden layer activations by a normal distribution. We show that it is possible to provide tighter guarantees for the suboptimality of MOMBO than the existing Monte Carlo sampling approaches. We also observe MOMBO to converge faster than these approaches in a large set of benchmark tasks.
Crop breeding is crucial in improving agricultural productivity while potentially decreasing land usage, greenhouse gas emissions, and water consumption. However, breeding programs are challenging due to long turnover times, high-dimensional decision spaces, long-term objectives, and the need to adapt to rapid climate change. This paper introduces the use of Reinforcement Learning (RL) to optimize simulated crop breeding programs. RL agents are trained to make optimal crop selection and cross-breeding decisions based on genetic information. To benchmark RL-based breeding algorithms, we introduce a suite of Gym environments. The study demonstrates the superiority of RL techniques over standard practices in terms of genetic gain when simulated in silico using real-world genomic maize data.
Off-policy actor-critic algorithms have shown promise in deep reinforcement learning for continuous control tasks. Their success largely stems from leveraging pessimistic state-action value function updates, which effectively address function approximation errors and improve performance. However, such pessimism can lead to under-exploration, constraining the agent's ability to explore/refine its policies. Conversely, optimism can counteract under-exploration, but it also carries the risk of excessive risk-taking and poor convergence if not properly balanced. Based on these insights, we introduce Utility Soft Actor-Critic (USAC), a novel framework within the actor-critic paradigm that enables independent control over the degree of pessimism/optimism for both the actor and the critic via interpretable parameters. USAC adapts its exploration strategy based on the uncertainty of critics through a utility function that allows us to balance between pessimism and optimism separately. By going beyond binary choices of optimism and pessimism, USAC represents a significant step towards achieving balance within off-policy actor-critic algorithms. Our experiments across various continuous control problems show that the degree of pessimism or optimism depends on the nature of the task. Furthermore, we demonstrate that USAC can outperform state-of-the-art algorithms for appropriately configured pessimism/optimism parameters.
Continuous action spaces in reinforcement learning (RL) are commonly defined as multidimensional intervals. While intervals usually reflect the action boundaries for tasks well, they can be challenging for learning because the typically large global action space leads to frequent exploration of irrelevant actions. Yet, little task knowledge can be sufficient to identify significantly smaller state-specific sets of relevant actions. Focusing learning on these relevant actions can significantly improve training efficiency and effectiveness. In this paper, we propose to focus learning on the set of relevant actions and introduce three continuous action masking methods for exactly mapping the action space to the state-dependent set of relevant actions. Thus, our methods ensure that only relevant actions are executed, enhancing the predictability of the RL agent and enabling its use in safety-critical applications. We further derive the implications of the proposed methods on the policy gradient. Using proximal policy optimization (PPO), we evaluate our methods on four control tasks, where the relevant action set is computed based on the system dynamics and a relevant state set. Our experiments show that the three action masking methods achieve higher final rewards and converge faster than the baseline without action masking.
Training autonomous agents with sparse rewards is a long-standing problem in online reinforcement learning (RL), due to low data efficiency. Prior work overcomes this challenge by extracting useful knowledge from offline data, often accomplished through the learning of action distribution from offline data and utilizing the learned distribution to facilitate online RL. However, since the offline data are given and fixed, the extracted knowledge is inherently limited, making it difficult to generalize to new tasks. We propose a novel approach that leverages offline data to learn a generative diffusion model, coined as Adaptive Trajectory Diffuser (ATraDiff). This model generates synthetic trajectories, serving as a form of data augmentation and consequently enhancing the performance of online RL methods. The key strength of our diffuser lies in its adaptability, allowing it to effectively handle varying trajectory lengths and mitigate distribution shifts between online and offline data. Because of its simplicity, ATraDiff seamlessly integrates with a wide spectrum of RL methods. Empirical evaluation shows that ATraDiff consistently achieves state-of-the-art performance across a variety of environments, with particularly pronounced improvements in complicated settings. Our code and demo video are available at https://atradiff.github.io .
This work studies the challenge of aligning large language models (LLMs) with offline preference data. We focus on alignment by Reinforcement Learning from Human Feedback (RLHF) in particular. While popular preference optimization methods exhibit good empirical performance in practice, they are not theoretically guaranteed to converge to the optimal policy and can provably fail when the data coverage is sparse by classical offline reinforcement learning (RL) results. On the other hand, a recent line of work has focused on theoretically motivated preference optimization methods with provable guarantees, but these are not computationally efficient for large-scale applications like LLM alignment. To bridge this gap, we propose SPAC, a new offline preference optimization method with self-play, inspired by the on-average pessimism technique from the offline RL literature, to be the first provable and scalable approach to LLM alignment. We both provide theoretical analysis for its convergence under single-policy concentrability for the general function approximation setting and demonstrate its competitive empirical performance for LLM alignment on a 7B Mistral model with Open LLM Leaderboard evaluations.
Many classic Reinforcement Learning (RL) algorithms rely on a Bellman operator, which involves an expectation over the next states, leading to the concept of bootstrapping. To introduce a form of pessimism, we propose to replace this expectation with an expectile. In practice, this can be very simply done by replacing the $L_2$ loss with a more general expectile loss for the critic. Introducing pessimism in RL is desirable for various reasons, such as tackling the overestimation problem (for which classic solutions are double Q-learning or the twin-critic approach of TD3) or robust RL (where transitions are adversarial). We study empirically these two cases. For the overestimation problem, we show that the proposed approach, ExpectRL, provides better results than a classic twin-critic. On robust RL benchmarks, involving changes of the environment, we show that our approach is more robust than classic RL algorithms. We also introduce a variation of ExpectRL combined with domain randomization which is competitive with state-of-the-art robust RL agents. Eventually, we also extend \ExpectRL with a mechanism for choosing automatically the expectile value, that is the degree of pessimism
Accuracy and timeliness are indeed often conflicting goals in prediction tasks. Premature predictions may yield a higher rate of false alarms, whereas delaying predictions to gather more information can render them too late to be useful. In applications such as wildfires, crimes, and traffic jams, timely forecasting are vital for safeguarding human life and property. Consequently, finding a balance between accuracy and timeliness is crucial. In this paper, we propose an early spatio-temporal forecasting model based on Multi-Objective reinforcement learning that can either implement an optimal policy given a preference or infer the preference based on a small number of samples. The model addresses two primary challenges: 1) enhancing the accuracy of early forecasting and 2) providing the optimal policy for determining the most suitable prediction time for each area. Our method demonstrates superior performance on three large-scale real-world datasets, surpassing existing methods in early spatio-temporal forecasting tasks.
Artificial agents' adaptability to novelty and alignment with intended behavior is crucial for their effective deployment. Reinforcement learning (RL) leverages novelty as a means of exploration, yet agents often struggle to handle novel situations, hindering generalization. To address these issues, we propose HackAtari, a framework introducing controlled novelty to the most common RL benchmark, the Atari Learning Environment. HackAtari allows us to create novel game scenarios (including simplification for curriculum learning), to swap the game elements' colors, as well as to introduce different reward signals for the agent. We demonstrate that current agents trained on the original environments include robustness failures, and evaluate HackAtari's efficacy in enhancing RL agents' robustness and aligning behavior through experiments using C51 and PPO. Overall, HackAtari can be used to improve the robustness of current and future RL algorithms, allowing Neuro-Symbolic RL, curriculum RL, causal RL, as well as LLM-driven RL. Our work underscores the significance of developing interpretable in RL agents.