Off-policy Reinforcement Learning Research Articles

Quadratic Tracking Control of Linear Stochastic Systems with Unknown Dynamics Using Average Off-Policy Q-Learning Method

This article investigates the optimal tracking control problem for data-based stochastic discrete-time linear systems. An average off-policy Q-learning algorithm is proposed to solve the optimal control problem with random disturbances. Compared with the existing off-policy reinforcement learning (RL) algorithm, the proposed average off-policy Q-learning algorithm avoids the assumption of an initial stability control. First, a pole placement strategy is used to design an initial stable control for systems with unknown dynamics. Second, the initial stable control is used to design a data-based average off-policy Q-learning algorithm. Then, this algorithm is used to solve the stochastic linear quadratic tracking (LQT) problem, and a convergence proof of the algorithm is provided. Finally, numerical examples show that this algorithm outperforms other algorithms in a simulation.

Heterogeneous unmanned swarm formation containment control based on reinforcement learning

In response to the problem of time-varying spherical formation control for a heterogeneous unmanned aerial vehicle (UAV) swarm system with dynamic uncertainty in the system model, this paper proposes an optimal distributed formation containment control method based on reinforcement learning. By combining the time-varying formation containment vector design with the distributed predefined time observer and constructing the augmented system of a multi-quadrotor UAV system and an observer, the problem of time-varying formation containment control in heterogeneous swarm systems is transformed into a stabilization problem. By introducing a value function with a discount factor, the stabilization problem of a heterogeneous UAV swarm system is transformed into an optimal control problem. Using the “actor-critic” neural network, combined with an off-policy reinforcement learning algorithm and distributed control methods, the solution to the formation containment controller is achieved in a data-driven manner. The stability of distributed observer and formation containment controller as well as the convergence of reinforcement learning algorithms are demonstrated using Lyapunov and related theory. The numerical simulation results demonstrate that the observation error of the designed distributed observer converges within a predefined time of 0.1 s, while the overall formation tracking error of the heterogeneous swarm converges within 2.74 s, thereby further validating the effectiveness and superiority of the designed control scheme.

Fully Data-Driven Robust Output Formation Tracking Control for Heterogeneous Multiagent System With Multiple Leaders and Actuator Faults.

This article investigates a fully data-driven method to solve the robust output formation tracking control problem for the multiagent system (MAS) under actuator faults. The outputs of the followers are controlled to track those of multiple leaders with respect to a convex point while achieving an expected time-varying formation. To obviate the requirement of various system prior knowledge in typical MAS control, a hierarchical frame is developed with three learning and control stages using the online measured data. First, a distributed adaptive observer is designed to coordinate the state convex of multiple leaders while estimating unknown dynamics. The adaptive mechanism relaxes the demand for global topology. Second, by collecting and reusing the online system data, an off-policy reinforcement learning (RL) method is proposed in a continuous form to acquire nominal feedback gains from partial observations of the followers. Essential system models are learned along with the RL process, while solutions to the output regulation equations are implicitly obtained. Third, a comprehensive robust controller is further presented based on the previous learning results. To address the actuator faults with efficiency loss and bias, the adaptive neural networks and robust compensations are utilized in a model-free manner. The output formation tracking is achieved under a derived feasibility condition while stabilities of the learning and control methods are analyzed. Finally, simulation results demonstrate the validity of this fully data-driven control frame.

Integrating an Ensemble Reward System into an Off-Policy Reinforcement Learning Algorithm for the Economic Dispatch of Small Modular Reactor-Based Energy Systems

Nuclear Integrated Energy Systems (NIES) have emerged as a comprehensive solution for navigating the changing energy landscape. They combine nuclear power plants with renewable energy sources, storage systems, and smart grid technologies to optimize energy production, distribution, and consumption across sectors, improving efficiency, reliability, and sustainability while addressing challenges associated with variability. The integration of Small Modular Reactors (SMRs) in NIES offers significant benefits over traditional nuclear facilities, although transferring involves overcoming legal and operational barriers, particularly in economic dispatch. This study proposes a novel off-policy Reinforcement Learning (RL) approach with an ensemble reward system to optimize economic dispatch for nuclear-powered generation companies equipped with an SMR, demonstrating superior accuracy and efficiency when compared to conventional methods and emphasizing RL’s potential to improve NIES profitability and sustainability. Finally, the research attempts to demonstrate the viability of implementing the proposed integrated RL approach in spot energy markets to maximize profits for nuclear-driven generation companies, establishing NIES’ profitability over competitors that rely on fossil fuel-based generation units to meet baseload requirements.

Mixed experience sampling for off-policy reinforcement learning

In deep reinforcement learning, experience replay is usually used to improve data efficiency and alleviate experience forgetting. However, online reinforcement learning is often influenced by the index of experience, which usually makes the phenomenon of unbalanced sampling. In addition, most experience replay methods ignore the differences among experiences, and cannot make full use of all experiences. Especially many “near”-policy experiences relatively relevant to the current policy are wasted, despite of the fact that they are beneficial for improving sample efficiency. This paper theoretically analyzes the influence of various factors on experience sampling, and then proposes a sampling method for experience replay based on frequency and similarity (FSER) to alleviate unbalanced sampling and increase the value of the sampled experiences. FSER prefers experiences that are rarely sampled or highly relevant to the current policy. FSER plays a critical role to balance the experience forgetting and wasting problems. Finally, FSER is combined with TD3 to achieve the state-of-the-art results in multiple tasks.

Integrating human learning and reinforcement learning: A novel approach to agent training

Off-policy reinforcement learning (RL) algorithms are known for improving sample efficiency by employing prior experiences in experience replay memory. However, most existing off-policy RL algorithms are bottlenecked by the slow convergence speed and demand of a large number of interaction samples. In contrast, on-policy RL algorithms converge fast and continuously generate new samples. However, their success heavily depends on the accuracy of the generated samples. To address these challenges, a novel RL framework called HI-FER (Human-learning Inspired Frequent Experience Replay) is proposed by mimicking the process of human learning. HI-FER employs a parallelized experience replay and repetitive training framework to expedite the convergence rate of off-policy algorithms, which imitates the function of the human brain’s parallel information processing and repetitive learning. Additionally, a periodic network reset strategy and dynamic memory updating are leveraged by imitating the forgetting mechanism of humans to prevent overfitting triggered by repetitive updating on limited experiences. Extensive comparison experiments and ablation studies are performed on benchmark environments to evaluate the proposed method. The empirical results demonstrate that HI-FER outperforms the baselines in terms of sample efficiency on state-based (14% improvements) and image-based (51% improvements) tasks from DMControl. Project website and code: https://github.com/Arya87/HI-FER.

Utilizing reinforcement learning for de novo drug design

AbstractDeep learning-based approaches for generating novel drug molecules with specific properties have gained a lot of interest in the last few years. Recent studies have demonstrated promising performance for string-based generation of novel molecules utilizing reinforcement learning. In this paper, we develop a unified framework for using reinforcement learning for de novo drug design, wherein we systematically study various on- and off-policy reinforcement learning algorithms and replay buffers to learn an RNN-based policy to generate novel molecules predicted to be active against the dopamine receptor DRD2. Our findings suggest that it is advantageous to use at least both top-scoring and low-scoring molecules for updating the policy when structural diversity is essential. Using all generated molecules at an iteration seems to enhance performance stability for on-policy algorithms. In addition, when replaying high, intermediate, and low-scoring molecules, off-policy algorithms display the potential of improving the structural diversity and number of active molecules generated, but possibly at the cost of a longer exploration phase. Our work provides an open-source framework enabling researchers to investigate various reinforcement learning methods for de novo drug design.

Optimal Tracking Control of Heterogeneous MASs Using Event-Driven Adaptive Observer and Reinforcement Learning.

This article considers the output tracking control problem of nonidentical linear multiagent systems (MASs) using a model-free reinforcement learning (RL) algorithm, where partial followers have no prior knowledge of the leader's information. To lower the communication and computing burden among agents, an event-driven adaptive distributed observer is proposed to predict the leader's system matrix and state, which consists of the estimated value of relative states governed by an edge-based predictor. Meanwhile, the integral input-based triggering condition is exploited to decide whether to transmit its private control input to its neighbors. Then, an RL-based state feedback controller for each agent is developed to solve the output tracking control problem, which is further converted into the optimal control problem by introducing a discounted performance function. Inhomogeneous algebraic Riccati equations (AREs) are derived to obtain the optimal solution of AREs. An off-policy RL algorithm is used to learn the solution of inhomogeneous AREs online without requiring any knowledge of the system dynamics. Rigorous analysis shows that under the proposed event-driven adaptive observer mechanism and RL algorithm, all followers are able to synchronize the leader's output asymptotically. Finally, a numerical simulation is demonstrated to verify the proposed approach in theory.

Colored Noise in PPO: Improved Exploration and Performance through Correlated Action Sampling

Proximal Policy Optimization (PPO), a popular on-policy deep reinforcement learning method, employs a stochastic policy for exploration. In this paper, we propose a colored noise-based stochastic policy variant of PPO. Previous research highlighted the importance of temporal correlation in action noise for effective exploration in off-policy reinforcement learning. Building on this, we investigate whether correlated noise can also enhance exploration in on-policy methods like PPO. We discovered that correlated noise for action selection improves learning performance and outperforms the currently popular uncorrelated white noise approach in on-policy methods. Unlike off-policy learning, where pink noise was found to be highly effective, we found that a colored noise, intermediate between white and pink, performed best for on-policy learning in PPO. We examined the impact of varying the amount of data collected for each update by modifying the number of parallel simulation environments for data collection and observed that with a larger number of parallel environments, more strongly correlated noise is beneficial. Due to the significant impact and ease of implementation, we recommend switching to correlated noise as the default noise source in PPO.

Double Buffers CEM-TD3: More Efficient Evolution and Richer Exploration

CEM-TD3 is a combination scheme using the simple cross-entropy method (CEM) and Twin Delayed Deep Deterministic policy gradient (TD3), and it achieves a satisfactory trade-off between performance and sample efficiency. However, we find that CEM-TD3 cannot fully address the low efficiency of policy search caused by CEM, and the policy gradient learning introduced by TD3 will weaken the diversity of individuals in the population. In this paper, we propose Double Buffers CEM-TD3 (DBCEM-TD3) that optimizes both CEM and TD3. For CEM, DBCEM-TD3 maintains an actor buffer to store the population required for evolution. In each iteration, it only needs to generate a small number of actors to replace the poor actors in the policy buffer to achieve more efficient evolution. The fitness of individuals in the actor buffer decreases exponentially with time, which can avoid premature convergence of the mean actor. For TD3, DBCEM-TD3 maintains a critic buffer with the same number of critics as the number of actors generated in each iteration, and each critic is trained independently by sampling from the shared replay buffer. In each iteration, each newly generated actor uses different critics to guide learning. This ensures more diverse behaviors among the learned actors, enabling richer experiences to be collected during the evaluation phase. We conduct experimental evaluations on five continuous control tasks provided by OpenAI Gym. DBCEM-TD3 outperforms CEM-TD3, TD3, and other classic off-policy reinforcement learning algorithms in terms of performance and sample efficiency.

Counterfactual Explanation for Fairness in Recommendation

Fairness-aware recommendation alleviates discrimination issues to build trustworthy recommendation systems. Explaining the causes of unfair recommendations is critical, as it promotes fairness diagnostics, and thus secures users’ trust in recommendation models. Existing fairness explanation methods suffer high computation burdens due to the large-scale search space and the greedy nature of the explanation search process. Besides, they perform feature-level optimizations with continuous values, which are not applicable to discrete attributes such as gender and age. In this work, we adopt counterfactual explanations from causal inference and propose to generate attribute-level counterfactual explanations, adapting to discrete attributes in recommendation models. We use real-world attributes from Heterogeneous Information Networks (HINs) to empower counterfactual reasoning on discrete attributes. We propose a Counterfactual Explanation for Fairness (CFairER) that generates attribute-level counterfactual explanations from HINs for item exposure fairness. Our CFairER conducts off-policy reinforcement learning to seek high-quality counterfactual explanations, with attentive action pruning reducing the search space of candidate counterfactuals. The counterfactual explanations help to provide rational and proximate explanations for model fairness, while the attentive action pruning narrows the search space of attributes. Extensive experiments demonstrate our proposed model can generate faithful explanations while maintaining favorable recommendation performance.

Robust hierarchical games of linear discrete-time systems based on off-policy model-free reinforcement learning

An off-policy model-free reinforcement learning (RL) algorithm is proposed for a robust hierarchical game while considering incomplete information and input constraints. The robust hierarchical game exhibits characteristics of a Stackelberg–Nash (SN) game, where equilibrium points are designated as Stackelberg–Nash–Saddle equilibrium (SNE) points. An off-policy method is employed for the RL algorithm, addressing input constraints by using excitation input instead of real-time update polices as control inputs. Moreover, a model-free method is implemented for the off-policy RL algorithm, accounting for the challenge posed by incomplete information. The goal of this paper is to develop an off-policy model-free RL algorithm to obtain approximate SNE polices of the robust hierarchical game with incomplete information and input constraints. Furthermore, the convergence and effectiveness of the off-policy model-free RL algorithm are guaranteed by proving the equivalence of Bellman equation between nominal SNE policies and approximate SNE policies. Finally, a simulation is provided to verify the advantage of the developed algorithm.

Distributed Minmax Strategy for Multiplayer Games: Stability, Robustness, and Algorithms.

This article studies a distributed minmax strategy for multiplayer games and develops reinforcement learning (RL) algorithms to solve it. The proposed minmax strategy is distributed, in the sense that it finds each player's optimal control policy without knowing all the other players' policies. Each player obtains its distributed control policy by solving a distributed algebraic Riccati equation in a multiplayer noncooperative game. This policy is found against the worst policies of all the other players. We guarantee the existence of distributed minmax solutions and study their L2 and asymptotic stabilities. Under mild conditions, the resulting minmax control policies are shown to improve robust gain and phase margins of multiplayer systems compared to the standard linear-quadratic regulator controller. Distributed minmax solutions are found using both model-based policy iteration and data-driven off-policy RL algorithms. Simulation examples verify the proposed formulation and its computational efficiency over the nondistributed Nash solutions.

Data-Efficient Off-Policy Learning for Distributed Optimal Tracking Control of HMAS With Unidentified Exosystem Dynamics.

In this article, a data-efficient off-policy reinforcement learning (RL) approach is proposed for distributed output tracking control of heterogeneous multiagent systems (HMASs) using approximate dynamic programming (ADP). Different from existing results that the kinematic model of the exosystem is addressable to partial or all agents, the dynamics of the exosystem are assumed to be completely unknown for all agents in this article. To solve this difficulty, an identifiable algorithm using the experience-replay method is designed for each agent to identify the system matrices of the novel reference model instead of the original exosystem. Then, an output-based distributed adaptive output observer is proposed to provide the estimations of the leader, and the proposed observer not only has a low dimension and less data transmission among agents but also is implemented in a fully distributed way. Besides, a data-efficient RL algorithm is given to design the optimal controller offline along with the system trajectories without solving output regulator equations. An ADP approach is developed to iteratively solve game algebraic Riccati equations (GAREs) using online information of state and input in an online way, which relaxes the requirement of knowing prior knowledge of agents' system matrices in an offline way. Finally, a numerical example is provided to verify the effectiveness of theoretical analysis.

Off-policy RL algorithms can be sample-efficient for continuous control via sample multiple reuse

Sample efficiency is one of the most critical issues for online reinforcement learning (RL). Existing methods achieve higher sample efficiency by adopting model-based methods, Q-ensemble, or better exploration mechanisms. We, instead, propose to train an off-policy RL agent via updating on a fixed sampled batch multiple times, thus reusing these samples and better exploiting them within a single optimization loop. We name our method sample multiple reuse (SMR). We theoretically show the properties of Q-learning with SMR, e.g., convergence. Furthermore, we incorporate SMR with off-the-shelf off-policy RL algorithms and conduct experiments on a variety of continuous control benchmarks. Empirical results show that SMR significantly boosts the sample efficiency of the base methods across most of the evaluated tasks without any hyperparameter tuning or additional tricks.

Re-attentive experience replay in off-policy reinforcement learning

Re-attentive experience replay in off-policy reinforcement learning

Data-Driven Inverse Reinforcement Learning Control for Linear Multiplayer Games.

This article proposes a data-driven inverse reinforcement learning (RL) control algorithm for nonzero-sum multiplayer games in linear continuous-time differential dynamical systems. The inverse RL problem in the games is solved by a learner reconstructing the unknown expert players' cost functions from demonstrated expert's optimal state and control input trajectories. The learner, thus, obtains the same control feedback gains and trajectories as the expert, only using data along system trajectories without knowing system dynamics. This article first proposes a model-based inverse RL policy iteration framework that has: 1) policy evaluation step for reconstructing cost matrices using Lyapunov functions; 2) state-reward weight improvement step using inverse optimal control (IOC); and 3) policy improvement step using optimal control. Based on the model-based policy iteration algorithm, this article further develops an online data-driven off-policy inverse RL algorithm without knowing any knowledge of system dynamics or expert control gains. Rigorous convergence and stability analysis of the algorithms are provided. It shows that the off-policy inverse RL algorithm guarantees unbiased solutions while probing noises are added to satisfy the persistence of excitation (PE) condition. Finally, two different simulation examples validate the effectiveness of the proposed algorithms.

Model-free control of underwater vehicle-manipulator system interacting with unknown environments

The underwater vehicle-manipulator (UVM) interaction tasks require that UVM can adapt to different environments and have good stability properties. However, the complicated structures of UVM and uncertainties of the environment lead to poor interaction. In addition, transient performance related to overshoot and convergence is required during the interaction to ensure the safety of the UVM and environment. In this paper, an inner-loop/outer-loop control formulation is developed for underwater vehicle-manipulator interacting with environment, subject to unknown system and environment dynamic. For the inner-loop, an approximation-free state feedback controller is designed for UVM system without a prior knowledge of dynamics, capable of guaranteeing trajectory tracking with prescribed performance. A prescribed function is constructed, which allows the convergence time and convergence accuracy to be predetermined. Moreover, it is independent of the initial conditions. For the outer-loop, an off-policy reinforcement learning algorithm is presented to optimize the parameters of the admittance model without requiring the information of environment dynamic. The proposed admittance optimization is feasible for arbitrary desired trajectory. Finally, simulation and experimental results validate the effectiveness of the proposed model-free tracking controller.

Intelligent optimization method for real-time decision-making in laminated cooling configurations through reinforcement learning

Optimizing conjugate heat transfer in laminated configurations is pivotal for efficiently cooling turbine blades. The traditional approaches have limitations in achieving rapid optimization when design conditions change. The emergence of reinforcement learning offers an opportunity to address this issue. However, there is currently a notable scarcity of research on the application of reinforcement learning in the optimization of laminated cooling configurations. This paper presents a novel method using the Deep Q-Network (DQN) model for real-time optimization of laminated cooling configurations. A dataset comprising 246 sets of historical design data was generated through conjugate simulations and augmented with the incorporation of physical information. After undergoing off-policy reinforcement learning, the intelligent model was employed to address optimization problems associated with temperature levels and coolant usage. Results indicate that the intelligent model achieves accurate and real-time optimization for laminated cooling configuration, with an average time of 3 ms. The ability to identify the correct optimization direction is attributed to the formulation of reward and penalty mechanisms in reinforcement learning. Real-time optimization is enabled by establishing a forward computational path from design conditions to optimization results. The findings of this study are easily transferrable and contribute to cost savings in design.

Autonomous navigation of mobile robots in unknown environments using off-policy reinforcement learning with curriculum learning

Reinforcement learning (RL) is effective for autonomous navigation tasks without prior knowledge of the environment. However, traditional mobile robot navigation algorithms, based on off-policy RL, often face challenges such as low sample efficiency during training and lack of adequate safety mechanisms. In this paper, we present an off-policy RL navigation model named Soft Actor-Critic with Curriculum Prioritization and Fuzzy Logic (SCF). The model uses energy as a prioritized evaluation metric for experience replay. And through task-level curriculum, the agent’s learning sequence is formulated, thereby enhancing sampling efficiency and safety. We propose a Curriculum-based Energy Prioritization (CEP) approach. It selects a replay trajectory that matches the current agent’s capability based on trajectory energy. Our results show that robots using off-policy RL often have limitations in dynamic obstacle avoidance. To rectify this, our model uses a fuzzy logic controller to enhance real-time obstacle avoidance. The SCF approach enables mobile robots to navigate adeptly in unpredictable and dynamic environments, ensuring optimal planning control while being safe and robust. Experiments in Gazebo simulation environment and real world confirm the effectiveness of our proposed method. The comparison results show the superior performance of this method, especially in unknown and dynamic environments.