Strategic Exploration in Reinforcement Learning - New Algorithms and Learning Guarantees

Reinforcement Learning (RL) Markov Decision Processes is studied with an emphasis on the well-studied exploration problem. We first formulate and discuss a definition of “efficient” algorithms that is termed Probably Approximately Correct (PAC) in RL. Next we provide general sufficient conditions for such an algorithm that applies to several different modeling assumptions. The conditions can be used to demonstrate that efficient learning is possible in finite MDPs, with either a model-based or model-free approach, in factored MDPs, and in continuous MDPs with linear dynamics. In the reinforcement-learning (RL) problem (Sutton & Barto 1998), an agent acts in an unknown or incompletely known environment with the goal of maximizing an external reward signal. In the most standard mathematical formulation of the problem, the environment is modeled as a Markov Decision Process (MDP) and the goal of the agent is to obtain near-optimal discounted return. Over the years, many algorithms have been proposed for this problem, but analyses of their performances have been relatively scarce. In fact, until recently, most theoretical guarantees have been that certain algorithms will discover an optimal policy in the limit, after an infinite amount of experience. In contrast, several attempts have been made to study “Probably Approximately Correct” or PAC-MDP algorithms, which exhibit near-optimal behavior in polynomial time and experience. This paper discusses several extensions of those results. We present a theorem that provides sufficient conditions for an algorithm to be PAC-MDP. We examine these conditions and show how they can be applied to prove that efficient learning is possible in three interesting scenerios: finite MDPs (i.e. the “Tabular case”),

Reinforcement learning with sparse rewards is still an open challenge. Classic methods rely on getting feedback via extrinsic rewards to train the agent, and in situations where this occurs very rarely the agent learns slowly or cannot learn at all. Similarly, if the agent receives also rewards that create suboptimal modes of the objective function, it will likely prematurely stop exploring. More recent methods add auxiliary intrinsic rewards to encourage exploration. However, auxiliary rewards lead to a non-stationary target for the Q-function. In this paper, we present a novel approach that (1) plans exploration actions far into the future by using a long-term visitation count, and (2) decouples exploration and exploitation by learning a separate function assessing the exploration value of the actions. Contrary to existing methods that use models of reward and dynamics, our approach is off-policy and model-free. We further propose new tabular environments for benchmarking exploration in reinforcement learning. Empirical results on classic and novel benchmarks show that the proposed approach outperforms existing methods in environments with sparse rewards, especially in the presence of rewards that create suboptimal modes of the objective function. Results also suggest that our approach scales gracefully with the size of the environment.

Reinforcement learning (RL) algorithms have been successfully applied to control tasks associated with unmanned aerial vehicles and robotics. In recent years, safe RL has been proposed to allow the safe execution of RL algorithms in industrial and mission-critical systems that operate in closed loops. However, if the system operating conditions change, such as when an unknown fault occurs, typical safe RL algorithms cannot adapt while retaining past knowledge. Continual RL algorithms have been proposed to address this issue. However, the impact of continual adaptation on the system’s safety is an understudied problem. In this paper, we study the intersection of safe and continual RL. First, we empirically demonstrate that a popular continual RL algorithm, elastic weight consolidation, does not satisfy safety constraints in nonlinear systems subject to varying operating conditions. Specifically, we study the MuJoCo HalfCheetah and Ant environments with velocity constraints and sudden joint loss non-stationarity. Then, we show that an agent trained using constrained policy optimization, a safe RL algorithm, experiences catastrophic forgetting in continual learning settings. With this in mind, we explore a simple reward-shaping method to ensure that elastic weight consolidation prioritizes remembering both safety and task performance for safety-constrained, nonlinear, and non-stationary systems.

P2P power trading based on reinforcement learning for nanogrid clusters

Bidirectional Human-Robot Learning: Imitation and Skill Improvement

Reinforcement learning is the problem of generating optimal behavior in a sequential decision-making environment given the opportunity of interacting with it. Many algorithms for solving reinforcem...

This paper proposes a reinforcement learning (RL) algorithm for the security problem of state estimation of cyber-physical system (CPS) under denial-of-service (DoS) attacks. The security of CPS will inevitably decline when faced with malicious cyber attacks. In order to analyze the impact of cyber attacks on CPS performance, a Kalman filter, as an adaptive state estimation technology, is combined with an RL method to evaluate the issue of system security, where estimation performance is adopted as an evaluation criterion. Then, the transition of estimation error covariance under a DoS attack is described as a Markov decision process, and the RL algorithm could be applied to resolve the optimal countermeasures. Meanwhile, the interactive combat between defender and attacker could be regarded as a two-player zero-sum game, where the Nash equilibrium policy exists but needs to be solved. Considering the energy constraints, the action selection of both sides will be restricted by setting certain cost functions. The proposed RL approach is designed from three different perspectives, including the defender, the attacker and the interactive game of two opposite sides. In addition, the framework of Q-learning and state–action–reward–state–action (SARSA) methods are investigated separately in this paper to analyze the influence of different RL algorithms. The results show that both algorithms obtain the corresponding optimal policy and the Nash equilibrium policy of the zero-sum interactive game. Through comparative analysis of two algorithms, it is verified that the differences between Q-Learning and SARSA could be applied effectively into the secure state estimation in CPS.

Recent advancement on reinforcement learning (RL) algorithms shows that effective learning of parametric action- selection policies can often be achieved through direct opti- mization of a performance lower bound subject to pre-defined policy behavioral constraints. Driven by this understanding, this paper seeks to develop new policy search techniques where RL is achieved through maximizing a performance lower bound ob- tained originally based on an Expectation-Maximization method. For reliable RL, our new learning techniques must also simul- taneously guarantee constrained policy behavioral changes mea- sured through KL divergence. Two separate approaches will be pursued to tackle our constrained policy optimization problems, resulting in two new RL algorithms. The first algorithm utilizes a conjugate gradient technique and a Bayesian learning method for approximate optimization. The second algorithm focuses on minimizing a loss function derived from solving the Lagrangian for constrained policy search. Both algorithms have been experi- mentally examined on several benchmark problems provided by OpenAI GYM. The experiment results clearly demonstrate that our algorithms can be highly effective in comparison to several well-known RL algorithms.

Reinforcement Learning (RL) has demonstrated significant potential in fields such as robotic control, autonomous driving, and financial decision-making. However, in complex environments, RL still faces challenges in stability and convergence. This study addresses this core issue by establishing a theoretical analysis framework that combines stochastic approximation theory and Lyapunov stability theory to rigorously analyze the convergence conditions and stability bounds of various RL algorithms in complex environments. Based on the theoretical analysis, we propose an Adaptive Stable RL (ASRL) algorithm, which employs dynamic regularized policy optimization and robust value function estimation to effectively mitigate policy oscillation and training divergence. Systematic experiments conducted in OpenAI Gym, MuJoCo, and customized non-stationary environments demonstrate that ASRL significantly outperforms baseline algorithms such as PPO and SAC in terms of convergence speed, final performance, and stability. Additionally, in industrial control and robotic navigation case studies, ASRL exhibits excellent adaptability and robustness. This research not only provides theoretical support for RL in complex environments but also offers optimization guidelines for algorithm design in practical applications.

This study introduces a novel controller based on a Reinforcement Learning (RL) algorithm for real-time adaptation of the stimulation pattern during FES-cycling. Core to our approach is the introduction of an RL agent that interacts with the cycling environment and learns through trial and error how to modulate the electrical charge applied to the stimulated muscle groups according to a predefined policy and while tracking a reference cadence. Instead of a static stimulation pattern to be modified by a control law, we hypothesized that a non-stationary baseline set of parameters would better adjust the amount of injected electrical charge to the time-varying characteristics of the musculature. Overground FES-assisted cycling sessions were performed by a subject with spinal cord injury (SCI AIS-A, T8). For tracking a predefined pedaling cadence, two closed-loop control laws were simultaneously used to modulate the pulse intensity of the stimulation channels responsible for evoking the muscle contractions. First, a Proportional-Integral (PI) controller was used to control the current amplitude of the stimulation channels over an initial parameter setting with predefined pulse amplitude, width and fixed frequency parameters. In parallel, an RL algorithm with a decayed-epsilon-greedy strategy was implemented to randomly explore nine different variations of pulse amplitude and width parameters over the same stimulation setting, aiming to adjust the injected electrical charge according to a predefined policy. The performance of this global control strategy was evaluated in two different RL settings and explored in two different cycling scenarios. The participant was able to pedal overground for distances over 3.5 km, and the results evidenced the RL agent learned to modify the stimulation pattern according to the predefined policy and was simultaneously able to track a predefined pedaling cadence. Despite the simplicity of our approach and the existence of more sophisticated RL algorithms, our method can be used to reduce the time needed to define stimulation patterns. Our results suggest interesting research possibilities to be explored in the future to improve cycling performance since more efficient stimulation cost dynamics can be explored and implemented for the agent to learn.

Recent research discovered that Reinforcement Learning (RL) algorithms with simple linear policies can achieve competitive performance as many state-of-the-art RL algorithms designed to train policies in the form of multi-layer neural networks. However, high learning performance is only achieved so far when policies are trained by jointly using multiple episodes of samples. An important research question remains as to whether linear policies can achieve cutting-edge performance when they are trained in a step-wise fashion (i.e., policies are iteratively updated based on every newly obtained sample). This paper presents a confirmatory answer to this question by developing a new RL algorithm called Primal-Dual Regular-gradient Actor-Critic (PD-RAC) as a generalization of RAC, which is a popular step-wise RL technique. Experiments on six benchmark control problems show that PD-RAC can achieve leading performance, in comparison to several recently developed baseline algorithms.

What is the next best drilling decision to make in a Field Development Plan (FDP)? This is the key question we address in this work. A FDP consists of a sequence of decisions. Each action we take affects the reservoir and conditions any future decision. The novelty of our proposed approach is the consideration of the sequential nature of the decisions through the framework of Dynamic Programming (DP) and Reinforcement Learning (RL). In this framework, each scheduled drilling decision depends on the observations acquired between drillings. This methodology allows moving the focus from a static Field Development Plan optimization to a more dynamic framework that we call Field Development Policy Optimization. In addition to the formulation of this new framework, we have applied this methodology to optimize the development of a real oil and gas field. We model the FDP optimization problem under subsurface uncertainty as a Partially Observable Markov Decision Process (POMDP) and solve it through a RL algorithm in order to find an optimal drilling policy. Our methodology works for a general reservoir with a given set of geological model representing. To speed up the learning process we utilize a trained Deep Recurrent Neural Network (RNN) to approximate the reservoir simulator flows, which are subsequently used to compute the economic performance of a drilling policy through its discounted cash flows. The RNN is trained and tested on a set of reservoir simulator runs over randomly sampled realizations of our reservoir model, well location, type and control sequences of a drilling plan. From all the possible decisions involved in a FDP, we focus here only on finding optimal adaptive well drilling schedules (locations of vertical wells and well types). The RL agent learns the best drilling schedule policy by generating simulated episodes of experience and iteratively improving the policy using a Q value function approximated by a neural network trained across episodes. The final solution consists of an adaptive Field Development Plan yielding the highest expected Net Present Value (NPV), computed within a given, available time budget. It specifies an adaptive drilling schedule of producer and injector, well locations and well controls, as a function of the information obtained at each drilling step. The methodology has been applied to an actual reservoir for infill well location decisions. In this case, our objective is finding the best well placement and well type for the next producer and injector wells, as well as the optimization of the control schedule for new and preexisting wells in the reservoir. Our results show the learning progress of our RL algorithm until finding the optimal drilling plan. The robustness of the solution is evaluated across the best-trained policies. Methodology and results have been validated using a brute force sampling approach. Both RL and brute force approaches were possible due to our fast-to-compute RNN approximation of the reservoir simulator. This work represents, to our knowledge, the first application of an end-to-end AI workflow for Field Development Policy Evaluation in real fields, based on Reinforcement Learning and Deep Learning. The proposed methodology puts together an optimal field evaluation in planning and a surveillance workflow for a reactive decision-making.

This paper investigates the error accumulation problem in centralized training and decentralized execution (CTDE) policy‐based multi‐agent reinforcement learning (MARL) algorithms, which arises from local observation inaccuracies. To address this issue, we propose a novel MARL algorithm that incorporates imitation learning using local observations. Firstly, by analysing the multi‐agent proximal policy optimization algorithm and examining the problems arising when global states are replaced with local observations, it is proved that insufficient observations can lead to information loss, thereby introducing errors of advantage function, and it is demonstrated that the generalized advantage estimation method accumulates errors during the training process. Then, imitation learning is introduced and a novel training framework that combines reinforcement learning and imitation learning is proposed. During the reinforcement learning phase, an MARL agent trained with global observations acts as an expert. Subsequently, imitation learning is applied to train another agent that mimics the expert's decisions using only local observations. Finally, the effectiveness of this algorithm is verified in some commonly used multi‐agent environments, which demonstrates its superior performance compared to traditional multi‐agent reinforcement learning algorithms.

Deep reinforcement learning (DRL) methods traditionally struggle with tasks where environment rewards are sparse or delayed, which entails that exploration remains one of the key challenges of DRL. Instead of solely relying on extrinsic rewards, many state-of-the-art methods use intrinsic curiosity as exploration signal. While they hold promise of better local exploration, discovering global exploration strategies is beyond the reach of current methods. We propose a novel end-to-end intrinsic reward formulation that introduces high-level exploration in reinforcement learning. Our curiosity signal is driven by a fast reward that deals with local exploration and a slow reward that incentivizes long-time horizon exploration strategies. We formulate curiosity as the error in an agent’s ability to reconstruct the observations given their contexts. Experimental results show that this high-level exploration enables our agents to outperform prior work in several Atari games.

Controlling feedback control systems in continuous action spaces has always been a challenging problem. Nevertheless, reinforcement learning is mainly an area of artificial intelligence (AI) because it has been used in process control for more than a decade. However, the existing algorithms are unable to provide satisfactory results. Therefore, this research uses a reinforcement learning (RL) algorithm to manage the control system. We propose an adaptive speed control of the motor system based on depth deterministic strategy gradient (DDPG). The actor-critic scenario using DDPG is implemented to build the RL agent. In addition, a framework has been created for traditional feedback control systems to make RL implementation easier for control systems. The RL algorithms are robust and proficient in using trial and error to search for the best strategy. Our proposed algorithm is a deep deterministic policy gradient, in which a large amount of training data trains the agent. Once the system is trained, the agent can automatically adjust the control parameters. The algorithm has been developed using Python 3.6 and the simulation results are evaluated in the MATLAB/Simulink environment. The performance of the proposed RL algorithm is compared with a proportional integral derivative (PID) controller and a linear quadratic regulator (LQR) controller. The simulation results of the proposed scheme are promising for the feedback control problems.