Safety is one of the key issues preventing the deployment of reinforcement learning techniques in real-world robots. While most approaches in the Safe Reinforcement Learning area do not require prior knowledge of constraints and robot kinematics and rely solely on data, it is often difficult to deploy them in complex real-world settings. Instead, model-based approaches that incorporate prior knowledge of the constraints and dynamics into the learning framework have proven capable of deploying the learning algorithm directly on the real robot. Unfortunately, while an approximated model of the robot dynamics is often available, the safety constraints are task-specific and hard to obtain: they may be too complicated to encode analytically, too expensive to compute, or it may be difficult to envision a priori the long-term safety requirements. In this paper, we bridge this gap by extending the safe exploration method, ATACOM, with learnable constraints, with a particular focus on ensuring long-term safety and handling of uncertainty. Our approach is competitive or superior to state-of-the-art methods in final performance while maintaining safer behavior during training.
The problem of safety for robotic systems has been extensively studied. However, little attention has been given to security issues for three-dimensional systems, such as quadrotors. Malicious adversaries can compromise robot sensors and communication networks, causing incidents, achieving illegal objectives, or even injuring people. This study first designs an intelligent control system for autonomous quadrotors. Then, it investigates the problems of optimal false data injection attack scheduling and countermeasure design for unmanned aerial vehicles. Using a state-of-the-art deep learning-based approach, an optimal false data injection attack scheme is proposed to deteriorate a quadrotor's tracking performance with limited attack energy. Subsequently, an optimal tracking control strategy is learned to mitigate attacks and recover the quadrotor's tracking performance. We base our work on Agilicious, a state-of-the-art quadrotor recently deployed for autonomous settings. This paper is the first in the United Kingdom to deploy this quadrotor and implement reinforcement learning on its platform. Therefore, to promote easy reproducibility with minimal engineering overhead, we further provide (1) a comprehensive breakdown of this quadrotor, including software stacks and hardware alternatives; (2) a detailed reinforcement-learning framework to train autonomous controllers on Agilicious agents; and (3) a new open-source environment that builds upon PyFlyt for future reinforcement learning research on Agilicious platforms. Both simulated and real-world experiments are conducted to show the effectiveness of the proposed frameworks in section 5.2.
This work proposes an approach that integrates reinforcement learning and model predictive control (MPC) to solve finite-horizon optimal control problems in mixed-logical dynamical systems efficiently. Optimization-based control of such systems with discrete and continuous decision variables entails the online solution of mixed-integer linear programs, which suffer from the curse of dimensionality. Our approach aims to mitigate this issue by decoupling the decision on the discrete variables from the decision on the continuous variables. In the proposed approach, reinforcement learning determines the discrete decision variables and simplifies the online optimization problem of the MPC controller from a mixed-integer linear program to a linear program, significantly reducing the computational time. A fundamental contribution of this work is the definition of the decoupled Q-function, which plays a crucial role in making the learning problem tractable in a combinatorial action space. We motivate the use of recurrent neural networks to approximate the decoupled Q-function and show how they can be employed in a reinforcement learning setting. Simulation experiments on a microgrid system using real-world data demonstrate that the proposed method substantially reduces the online computation time of MPC while maintaining high feasibility and low suboptimality.
The Job Shop Scheduling Problem (JSSP) is a complex combinatorial optimization problem. While online Reinforcement Learning (RL) has shown promise by quickly finding acceptable solutions for JSSP, it faces key limitations: it requires extensive training interactions from scratch leading to sample inefficiency, cannot leverage existing high-quality solutions, and often yields suboptimal results compared to traditional methods like Constraint Programming (CP). We introduce Offline Reinforcement Learning for Learning to Dispatch (Offline-LD), which addresses these limitations by learning from previously generated solutions. Our approach is motivated by scenarios where historical scheduling data and expert solutions are available, although our current evaluation focuses on benchmark problems. Offline-LD adapts two CQL-based Q-learning methods (mQRDQN and discrete mSAC) for maskable action spaces, introduces a novel entropy bonus modification for discrete SAC, and exploits reward normalization through preprocessing. Our experiments demonstrate that Offline-LD outperforms online RL on both generated and benchmark instances. Notably, by introducing noise into the expert dataset, we achieve similar or better results than those obtained from the expert dataset, suggesting that a more diverse training set is preferable because it contains counterfactual information.
In this research paper, we investigate into a paper named "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem" [arXiv:1706.10059]. It is a portfolio management problem which is solved by deep learning techniques. The original paper proposes a financial-model-free reinforcement learning framework, which consists of the Ensemble of Identical Independent Evaluators (EIIE) topology, a Portfolio-Vector Memory (PVM), an Online Stochastic Batch Learning (OSBL) scheme, and a fully exploiting and explicit reward function. Three different instants are used to realize this framework, namely a Convolutional Neural Network (CNN), a basic Recurrent Neural Network (RNN), and a Long Short-Term Memory (LSTM). The performance is then examined by comparing to a number of recently reviewed or published portfolio-selection strategies. We have successfully replicated their implementations and evaluations. Besides, we further apply this framework in the stock market, instead of the cryptocurrency market that the original paper uses. The experiment in the cryptocurrency market is consistent with the original paper, which achieve superior returns. But it doesn't perform as well when applied in the stock market.
In a reinforcement learning (RL) setting, the agent's optimal strategy heavily depends on her risk preferences and the underlying model dynamics of the training environment. These two aspects influence the agent's ability to make well-informed and time-consistent decisions when facing testing environments. In this work, we devise a framework to solve robust risk-aware RL problems where we simultaneously account for environmental uncertainty and risk with a class of dynamic robust distortion risk measures. Robustness is introduced by considering all models within a Wasserstein ball around a reference model. We estimate such dynamic robust risk measures using neural networks by making use of strictly consistent scoring functions, derive policy gradient formulae using the quantile representation of distortion risk measures, and construct an actor-critic algorithm to solve this class of robust risk-aware RL problems. We demonstrate the performance of our algorithm on a portfolio allocation example.
Reinforcement learning (RL) policies may exhibit unsafe behavior and are hard to explain. We use counterfactual large language model reasoning to enhance RL policy safety post-training. We show that our approach improves and helps to explain the RL policy safety.
Reinforcement learning from human feedback (RLHF) has become a key method for aligning large language models (LLMs) with human preferences through the use of reward models. However, traditional reward models typically generate point estimates, which oversimplify the diversity and complexity of human values and preferences. In this paper, we introduce Quantile Reward Models (QRMs), a novel approach to reward modeling that learns a distribution over rewards instead of a single scalar value. Our method uses quantile regression to estimate a full, potentially multimodal distribution over preferences, providing a more powerful and nuanced representation of preferences. This distributional approach can better capture the diversity of human values, addresses label noise, and accommodates conflicting preferences by modeling them as distinct modes in the distribution. Our experimental results show that QRM outperforms comparable traditional point-estimate models on RewardBench. Furthermore, we demonstrate that the additional information provided by the distributional estimates can be utilized in downstream applications, such as risk-aware reinforcement learning, resulting in LLM policies that generate fewer extremely negative responses. Our code and model are released at https://github.com/Nicolinho/QRM.
Although deep reinforcement learning (DRL) approaches in audio signal processing have seen substantial progress in recent years, audio-driven DRL for tasks such as navigation, gaze control and head-orientation control in the context of human-robot interaction have received little attention. Here, we propose an audio-driven DRL framework in which we utilise deep Q-learning to develop an autonomous agent that orients towards a talker in the acoustic environment based on stereo speech recordings. Our results show that the agent learned to perform the task at a near perfect level when trained on speech segments in anechoic environments (that is, without reverberation). The presence of reverberation in naturalistic acoustic environments affected the agent's performance, although the agent still substantially outperformed a baseline, randomly acting agent. Finally, we quantified the degree of generalization of the proposed DRL approach across naturalistic acoustic environments. Our experiments revealed that policies learned by agents trained on medium or high reverb environments generalized to low reverb environments, but policies learned by agents trained on anechoic or low reverb environments did not generalize to medium or high reverb environments. Taken together, this study demonstrates the potential of audio-driven DRL for tasks such as head-orientation control and highlights the need for training strategies that enable robust generalization across environments for real-world audio-driven DRL applications.
Graph path search is a classic computer science problem that has been recently approached with Reinforcement Learning (RL) due to its potential to outperform prior methods. Existing RL techniques typically assume a global view of the network, which is not suitable for large-scale, dynamic, and privacy-sensitive settings. An area of particular interest is search in social networks due to its numerous applications. Inspired by seminal work in experimental sociology, which showed that decentralized yet efficient search is possible in social networks, we frame the problem as a collaborative task between multiple agents equipped with a limited local view of the network. We propose a multi-agent approach for graph path search that successfully leverages both homophily and structural heterogeneity. Our experiments, carried out over synthetic and real-world social networks, demonstrate that our model significantly outperforms learned and heuristic baselines. Furthermore, our results show that meaningful embeddings for graph navigation can be constructed using reward-driven learning.
Deep Reinforcement Learning techniques are achieving state-of-the-art results in robust legged locomotion. While there exists a wide variety of legged platforms such as quadruped, humanoids, and hexapods, the field is still missing a single learning framework that can control all these different embodiments easily and effectively and possibly transfer, zero or few-shot, to unseen robot embodiments. We introduce URMA, the Unified Robot Morphology Architecture, to close this gap. Our framework brings the end-to-end Multi-Task Reinforcement Learning approach to the realm of legged robots, enabling the learned policy to control any type of robot morphology. The key idea of our method is to allow the network to learn an abstract locomotion controller that can be seamlessly shared between embodiments thanks to our morphology-agnostic encoders and decoders. This flexible architecture can be seen as a potential first step in building a foundation model for legged robot locomotion. Our experiments show that URMA can learn a locomotion policy on multiple embodiments that can be easily transferred to unseen robot platforms in simulation and the real world.
Reinforcement Learning (RL) is a learning paradigm in which the agent learns from its environment through trial and error. Deep reinforcement learning (DRL) algorithms represent the agent's policies using neural networks, making their decisions difficult to interpret. Explaining the behaviour of DRL agents is necessary to advance user trust, increase engagement, and facilitate integration with real-life tasks. Semifactual explanations aim to explain an outcome by providing "even if" scenarios, such as "even if the car were moving twice as slowly, it would still have to swerve to avoid crashing". Semifactuals help users understand the effects of different factors on the outcome and support the optimisation of resources. While extensively studied in psychology and even utilised in supervised learning, semifactuals have not been used to explain the decisions of RL systems. In this work, we develop a first approach to generating semifactual explanations for RL agents. We start by defining five properties of desirable semifactual explanations in RL and then introducing SGRL-Rewind and SGRL-Advance, the first algorithms for generating semifactual explanations in RL. We evaluate the algorithms in two standard RL environments and find that they generate semifactuals that are easier to reach, represent the agent's policy better, and are more diverse compared to baselines. Lastly, we conduct and analyse a user study to assess the participant's perception of semifactual explanations of the agent's actions.
This document contains an educational introduction to the problem of sparsifying parametric models with L0 regularization. We utilize this approach together with dictionary learning to learn sparse polynomial policies for deep reinforcement learning to control parametric partial differential equations. The code and a tutorial are provided here: https://github.com/nicob15/Sparsifying-Parametric-Models-with-L0.
Reinforcement Learning (RL) is highly dependent on the meticulous design of the reward function. However, accurately assigning rewards to each state-action pair in Long-Term RL (LTRL) challenges is formidable. Consequently, RL agents are predominantly trained with expert guidance. Drawing on the principles of ordinal utility theory from economics, we propose a novel reward estimation algorithm: ELO-Rating based RL (ERRL). This approach is distinguished by two main features. Firstly, it leverages expert preferences over trajectories instead of cardinal rewards (utilities) to compute the ELO rating of each trajectory as its reward. Secondly, a new reward redistribution algorithm is introduced to mitigate training volatility in the absence of a fixed anchor reward. Our method demonstrates superior performance over several leading baselines in long-term scenarios (extending up to 5000 steps), where conventional RL algorithms falter. Furthermore, we conduct a thorough analysis of how expert preferences affect the outcomes.
In the following report, we describe the solution we propose for the AI Olympics competition held at IROS 2024. Our solution is based on a Model-free Deep Reinforcement Learning approach combined with an evolutionary strategy. We will briefly describe the algorithms that have been used and then provide details of the approach
In the field of autonomous driving, developing safe and trustworthy autonomous driving policies remains a significant challenge. Recently, Reinforcement Learning with Human Feedback (RLHF) has attracted substantial attention due to its potential to enhance training safety and sampling efficiency. Nevertheless, existing RLHF-enabled methods often falter when faced with imperfect human demonstrations, potentially leading to training oscillations or even worse performance than rule-based approaches. Inspired by the human learning process, we propose Physics-enhanced Reinforcement Learning with Human Feedback (PE-RLHF). This novel framework synergistically integrates human feedback (e.g., human intervention and demonstration) and physics knowledge (e.g., traffic flow model) into the training loop of reinforcement learning. The key advantage of PE-RLHF is its guarantee that the learned policy will perform at least as well as the given physics-based policy, even when human feedback quality deteriorates, thus ensuring trustworthy safety improvements. PE-RLHF introduces a Physics-enhanced Human-AI (PE-HAI) collaborative paradigm for dynamic action selection between human and physics-based actions, employs a reward-free approach with a proxy value function to capture human preferences, and incorporates a minimal intervention mechanism to reduce the cognitive load on human mentors. Extensive experiments across diverse driving scenarios demonstrate that PE-RLHF significantly outperforms traditional methods, achieving state-of-the-art (SOTA) performance in safety, efficiency, and generalizability, even with varying quality of human feedback. The philosophy behind PE-RLHF not only advances autonomous driving technology but can also offer valuable insights for other safety-critical domains. Demo video and code are available at: \https://zilin-huang.github.io/PE-RLHF-website/
Recurrent Neural Networks (RNNs) are used to learn representations in partially observable environments. For agents that learn online and continually interact with the environment, it is desirable to train RNNs with real-time recurrent learning (RTRL); unfortunately, RTRL is prohibitively expensive for standard RNNs. A promising direction is to use linear recurrent architectures (LRUs), where dense recurrent weights are replaced with a complex-valued diagonal, making RTRL efficient. In this work, we build on these insights to provide a lightweight but effective approach for training RNNs in online RL. We introduce Recurrent Trace Units (RTUs), a small modification on LRUs that we nonetheless find to have significant performance benefits over LRUs when trained with RTRL. We find RTUs significantly outperform other recurrent architectures across several partially observable environments while using significantly less computation.
This work proposes a control-informed reinforcement learning (CIRL) framework that integrates proportional-integral-derivative (PID) control components into the architecture of deep reinforcement learning (RL) policies. The proposed approach augments deep RL agents with a PID controller layer, incorporating prior knowledge from control theory into the learning process. CIRL improves performance and robustness by combining the best of both worlds: the disturbance-rejection and setpoint-tracking capabilities of PID control and the nonlinear modeling capacity of deep RL. Simulation studies conducted on a continuously stirred tank reactor system demonstrate the improved performance of CIRL compared to both conventional model-free deep RL and static PID controllers. CIRL exhibits better setpoint-tracking ability, particularly when generalizing to trajectories outside the training distribution, suggesting enhanced generalization capabilities. Furthermore, the embedded prior control knowledge within the CIRL policy improves its robustness to unobserved system disturbances. The control-informed RL framework combines the strengths of classical control and reinforcement learning to develop sample-efficient and robust deep reinforcement learning algorithms, with potential applications in complex industrial systems.
The escalating prevalence of cannabis use poses a significant public health challenge globally. In the U.S., cannabis use is more prevalent among emerging adults (EAs) (ages 18-25) than any other age group, with legalization in the multiple states contributing to a public perception that cannabis is less risky than in prior decades. To address this growing concern, we developed MiWaves, a reinforcement learning (RL) algorithm designed to optimize the delivery of personalized intervention prompts to reduce cannabis use among EAs. MiWaves leverages domain expertise and prior data to tailor the likelihood of delivery of intervention messages. This paper presents a comprehensive overview of the algorithm's design, including key decisions and experimental outcomes. The finalized MiWaves RL algorithm was deployed in a clinical trial from March to May 2024.
Centralized Training for Decentralized Execution where agents are trained offline in a centralized fashion and execute online in a decentralized manner, has become a popular approach in Multi-Agent Reinforcement Learning (MARL). In particular, it has become popular to develop actor-critic methods that train decentralized actors with a centralized critic where the centralized critic is allowed access global information of the entire system, including the true system state. Such centralized critics are possible given offline information and are not used for online execution. While these methods perform well in a number of domains and have become a de facto standard in MARL, using a centralized critic in this context has yet to be sufficiently analyzed theoretically or empirically. In this paper, we therefore formally analyze centralized and decentralized critic approaches, and analyze the effect of using state-based critics in partially observable environments. We derive theories contrary to the common intuition: critic centralization is not strictly beneficial, and using state values can be harmful. We further prove that, in particular, state-based critics can introduce unexpected bias and variance compared to history-based critics. Finally, we demonstrate how the theory applies in practice by comparing different forms of critics on a wide range of common multi-agent benchmarks. The experiments show practical issues such as the difficulty of representation learning with partial observability, which highlights why the theoretical problems are often overlooked in the literature.