在加强型学习系统中用伪熵进行不确定性估计

加强型学习系统是一种与没有约束的，未知的环境相互作用的系统，学习系统的目标在大最大可能地获取累积奖励信号，这个奖励信号在有限，未知的生命周期由系统所处的环境中得到，对于一个加强型学习系统，困难之一在于奖励信号非常稀疏，尤其是对于只有时延信号的系统，已有的加强型学习方法以价值函数的形式贮存奖励信号，例如著名的Ｑ－学习。本文提出了一个基于状态的不生估计模型的方法，这个算法对有利用存贮于价值函数中的奖励     

MOSAIC for multiple-reward environments

Reinforcement learning (RL) can provide a basic framework for autonomous robots to learn to control and maximize future cumulative rewards in complex environments. To achieve high performance, RL controllers must consider the complex external dynamics for movements and task (reward function) and optimize control commands. For example, a robot playing tennis and squash needs to cope with the different dynamics of a tennis or squash racket and such dynamic environmental factors as the wind. In addition, this robot has to tailor its tactics simultaneously under the rules of either game. This double complexity of the external dynamics and reward function sometimes becomes more complex when both the multiple dynamics and multiple reward functions switch implicitly, as in the situation of a real (multi-agent) game of tennis where one player cannot observe the intention of her opponents or her partner. The robot must consider its opponent's and its partner's unobservable behavioral goals (reward function). In this article, we address how an RL agent should be designed to handle such double complexity of dynamics and reward. We have previously proposed modular selection and identification for control (MOSAIC) to cope with nonstationary dynamics where appropriate controllers are selected and learned among many candidates based on the error of its paired dynamics predictor: the forward model. Here we extend this framework for RL and propose MOSAIC-MR architecture. It resembles MOSAIC in spirit and selects and learns an appropriate RL controller based on the RL controller's TD error using the errors of the dynamics (the forward model) and the reward predictors. Furthermore, unlike other MOSAIC variants for RL, RL controllers are not a priori paired with the fixed predictors of dynamics and rewards. The simulation results demonstrate that MOSAIC-MR outperforms other counterparts because of this flexible association ability among RL controllers, forward models, and reward predictors.     

Bayes-adaptive hierarchical MDPs

Reinforcement learning (RL) is an area of machine learning that is concerned with how an agent learns to make decisions sequentially in order to optimize a particular performance measure. For achieving such a goal, the agent has to choose either 1) exploiting previously known knowledge that might end up at local optimality or 2) exploring to gather new knowledge that expects to improve the current performance. Among other RL algorithms, Bayesian model-based RL (BRL) is well-known to be able to trade-off between exploitation and exploration optimally via belief planning, i.e. partially observable Markov decision process (POMDP). However, solving that POMDP often suffers from curse of dimensionality and curse of history. In this paper, we make two major contributions which are: 1) an integration framework of temporal abstraction into BRL that eventually results in a hierarchical POMDP formulation, which can be solved online using a hierarchical sample-based planning solver; 2) a subgoal discovery method for hierarchical BRL that automatically discovers useful macro actions to accelerate learning. In the experiment section, we demonstrate that the proposed approach can scale up to much larger problems. On the other hand, the agent is able to discover useful subgoals for speeding up Bayesian reinforcement learning.     

强化学习算法中启发式回报函数的设计及其收敛性分析

(中国科学院沈阳自动化所机器人学重点实验室沈阳110016)     

Model-based average reward reinforcement learning

Reinforcement Learning (RL) is the study of programs that improve their performance by receiving rewards and punishments from the environment. Most RL methods optimize the discounted total reward received by an agent, while, in many domains, the natural criterion is to optimize the average reward per time step. In this paper, we introduce a model-based Averagereward Reinforcement Learning method called H-learning and show that it converges more quickly and robustly than its discounted counterpart in the domain of scheduling a simulated Automatic Guided Vehicle (AGV). We also introduce a version of H-learning that automatically explores the unexplored parts of the state space, while always choosing greedy actions with respect to the current value function. We show that this “Auto-exploratory H-Learning” performs better than the previously studied exploration strategies. To scale H-learning to larger state spaces, we extend it to learn action models and reward functions in the form of dynamic Bayesian networks, and approximate its value function using local linear regression. We show that both of these extensions are effective in significantly reducing the space requirement of H-learning and making it converge faster in some AGV scheduling tasks.     

Integration of imitation learning using GAIL and reinforcement learning using task-achievement rewards via probabilistic graphical model

The integration of reinforcement learning (RL) and imitation learning (IL) is an important problem that has long been studied in the field of intelligent robotics. RL optimizes policies to maximize the cumulative reward, whereas IL attempts to extract general knowledge about the trajectories demonstrated by experts, i.e, demonstrators. Because each has its own drawbacks, many methods combining them and compensating for each set of drawbacks have been explored thus far. However, many of these methods are heuristic and do not have a solid theoretical basis. This paper presents a new theory for integrating RL and IL by extending the probabilistic graphical model (PGM) framework for RL, control as inference. We develop a new PGM for RL with multiple types of rewards, called probabilistic graphical model for Markov decision processes with multiple optimality emissions (pMDP-MO). Furthermore, we demonstrate that the integrated learning method of RL and IL can be formulated as a probabilistic inference of policies on pMDP-MO by considering the discriminator in generative adversarial imitation learning (GAIL) as an additional optimality emission. We adapt the GAIL and task-achievement reward to our proposed framework, achieving significantly better performance than policies trained with baseline methods.     

Exploiting submodular value functions for scaling up active perception

In active perception tasks, an agent aims to select sensory actions that reduce its uncertainty about one or more hidden variables. For example, a mobile robot takes sensory actions to efficiently navigate in a new environment. While partially observable Markov decision processes (POMDPs) provide a natural model for such problems, reward functions that directly penalize uncertainty in the agent’s belief can remove the piecewise-linear and convex (PWLC) property of the value function required by most POMDP planners. Furthermore, as the number of sensors available to the agent grows, the computational cost of POMDP planning grows exponentially with it, making POMDP planning infeasible with traditional methods. In this article, we address a twofold challenge of modeling and planning for active perception tasks. We analyze \(\rho \)POMDP and POMDP-IR, two frameworks for modeling active perception tasks, that restore the PWLC property of the value function. We show the mathematical equivalence of these two frameworks by showing that given a \(\rho \)POMDP along with a policy, they can be reduced to a POMDP-IR and an equivalent policy (and vice-versa). We prove that the value function for the given \(\rho \)POMDP (and the given policy) and the reduced POMDP-IR (and the reduced policy) is the same. To efficiently plan for active perception tasks, we identify and exploit the independence properties of POMDP-IR to reduce the computational cost of solving POMDP-IR (and \(\rho \)POMDP). We propose greedy point-based value iteration (PBVI), a new POMDP planning method that uses greedy maximization to greatly improve scalability in the action space of an active perception POMDP. Furthermore, we show that, under certain conditions, including submodularity, the value function computed using greedy PBVI is guaranteed to have bounded error with respect to the optimal value function. We establish the conditions under which the value function of an active perception POMDP is guaranteed to be submodular. Finally, we present a detailed empirical analysis on a dataset collected from a multi-camera tracking system employed in a shopping mall. Our method achieves similar performance to existing methods but at a fraction of the computational cost leading to better scalability for solving active perception tasks.     

Robust reinforcement learning

This letter proposes a new reinforcement learning (RL) paradigm that explicitly takes into account input disturbance as well as modeling errors. The use of environmental models in RL is quite popular for both offline learning using simulations and for online action planning. However, the difference between the model and the real environment can lead to unpredictable, and often unwanted, results. Based on the theory of H(infinity) control, we consider a differential game in which a "disturbing" agent tries to make the worst possible disturbance while a "control" agent tries to make the best control input. The problem is formulated as finding a min-max solution of a value function that takes into account the amount of the reward and the norm of the disturbance. We derive online learning algorithms for estimating the value function and for calculating the worst disturbance and the best control in reference to the value function. We tested the paradigm, which we call robust reinforcement learning (RRL), on the control task of an inverted pendulum. In the linear domain, the policy and the value function learned by online algorithms coincided with those derived analytically by the linear H(infinity) control theory. For a fully nonlinear swing-up task, RRL achieved robust performance with changes in the pendulum weight and friction, while a standard reinforcement learning algorithm could not deal with these changes. We also applied RRL to the cart-pole swing-up task, and a robust swing-up policy was acquired.     

Reinforcement learning in spacecraft control applications: Advances,prospects, and challenges

This paper presents and analyzes Reinforcement Learning (RL) based approaches to solve spacecraft control problems. Different application fields are considered, e.g., guidance, navigation and control systems for spacecraft landing on celestial bodies, constellation orbital control, and maneuver planning in orbit transfers. It is discussed how RL solutions can address the emerging needs of designing spacecraft with highly autonomous on-board capabilities and implementing controllers (i.e., RL agents) robust to system uncertainties and adaptive to changing environments. For each application field, the RL framework core elements (e.g., the reward function, the RL algorithm and the environment model used for the RL agent training) are discussed with the aim of providing some guidelines in the formulation of spacecraft control problems via a RL framework. At the same time, the adoption of RL in real space projects is also analyzed. Different open points are identified and discussed, e.g., the availability of high-fidelity simulators for the RL agent training and the verification of RL-based solutions. This way, recommendations for future work are proposed with the aim of reducing the technological gap between the solutions proposed by the academic community and the needs/requirements of the space industry.     

A Bayesian exploration-exploitation approach for optimal online sensing and planning with a visually guided mobile robot

We address the problem of online path planning for optimal sensing with a mobile robot. The objective of the robot is to learn the most about its pose and the environment given time constraints. We use a POMDP with a utility function that depends on the belief state to model the finite horizon planning problem. We replan as the robot progresses throughout the environment. The POMDP is high-dimensional, continuous, non-differentiable, nonlinear, non-Gaussian and must be solved in real-time. Most existing techniques for stochastic planning and reinforcement learning are therefore inapplicable. To solve this extremely complex problem, we propose a Bayesian optimization method that dynamically trades off exploration (minimizing uncertainty in unknown parts of the policy space) and exploitation (capitalizing on the current best solution). We demonstrate our approach with a visually-guide mobile robot. The solution proposed here is also applicable to other closely-related domains, including active vision, sequential experimental design, dynamic sensing and calibration with mobile sensors.     

Planning to see: A hierarchical approach to planning visual actions on a robot using POMDPs

Flexible, general-purpose robots need to autonomously tailor their sensing and information processing to the task at hand. We pose this challenge as the task of planning under uncertainty. In our domain, the goal is to plan a sequence of visual operators to apply on regions of interest (ROIs) in images of a scene, so that a human and a robot can jointly manipulate and converse about objects on a tabletop. We pose visual processing management as an instance of probabilistic sequential decision making, and specifically as a Partially Observable Markov Decision Process (POMDP). The POMDP formulation uses models that quantitatively capture the unreliability of the operators and enable a robot to reason precisely about the trade-offs between plan reliability and plan execution time. Since planning in practical-sized POMDPs is intractable, we partially ameliorate this intractability for visual processing by defining a novel hierarchical POMDP based on the cognitive requirements of the corresponding planning task. We compare our hierarchical POMDP planning system (HiPPo) with a non-hierarchical POMDP formulation and the Continual Planning (CP) framework that handles uncertainty in a qualitative manner. We show empirically that HiPPo and CP outperform the naive application of all visual operators on all ROIs. The key result is that the POMDP methods produce more robust plans than CP or the naive visual processing. In summary, visual processing problems represent a challenging domain for planning techniques and our hierarchical POMDP-based approach for visual processing management opens up a promising new line of research.     

Transfer in variable-reward hierarchical reinforcement learning

Transfer learning seeks to leverage previously learned tasks to achieve faster learning in a new task. In this paper, we consider transfer learning in the context of related but distinct Reinforcement Learning (RL) problems. In particular, our RL problems are derived from Semi-Markov Decision Processes (SMDPs) that share the same transition dynamics but have different reward functions that are linear in a set of reward features. We formally define the transfer learning problem in the context of RL as learning an efficient algorithm to solve any SMDP drawn from a fixed distribution after experiencing a finite number of them. Furthermore, we introduce an online algorithm to solve this problem, Variable-Reward Reinforcement Learning (VRRL), that compactly stores the optimal value functions for several SMDPs, and uses them to optimally initialize the value function for a new SMDP. We generalize our method to a hierarchical RL setting where the different SMDPs share the same task hierarchy. Our experimental results in a simplified real-time strategy domain show that significant transfer learning occurs in both flat and hierarchical settings. Transfer is especially effective in the hierarchical setting where the overall value functions are decomposed into subtask value functions which are more widely amenable to transfer across different SMDPs.     

一种动态不确定环境中机器人路径规划方法

针对动态不确定环境下的机器人路径规划问题,将部分可观察马尔可夫决策过程（POMDP）与人工势场法（APF）的优点相结合,提出一种新的机器人路径规划方法。该方法充分考虑了实际环境中信息的部分可观测性,并且利用APF无需大量计算的优点指导POMDP算法的奖赏值设定,以提高POMDP算法的决策效率。仿真实验表明,所提出的算法拥有较高的搜索效率,能够快速地到达目标点。     

基于平均序列累计奖赏的自适应ε-greedy策略

探索与利用的权衡是强化学习的挑战之一。探索使智能体为进一步改进策略而采取新的动作,而利用使智能体采用历史经验中的信息以最大化累计奖赏。深度强化学习中常用“[ε]-greedy”策略处理探索与利用的权衡问题,未考虑影响智能体做出决策的其他因素,具有一定的盲目性。针对此问题提出一种自适应调节探索因子的[ε]-greedy策略,该策略依据智能体每完成一次任务所获得的序列累计奖赏值指导智能体进行合理的探索或利用。序列累计奖赏值越大,说明当前智能体所采用的有效动作越多,减小探索因子以便更多地利用历史经验。反之,序列累计奖赏值越小,说明当前策略还有改进的空间,增大探索因子以便探索更多可能的动作。实验结果证明改进的策略在Playing Atari 2600视频游戏中取得了更高的平均奖赏值,说明改进的策略能更好地权衡探索与利用。     

Software Agent with Reinforcement Learning Approach for Medical Image Segmentation

Many image segmentation solutions are problem-based. Medical images have very similar grey level and texture among the interested objects. Therefore, medical image segmentation requires improvements although there have been researches done since the last few decades. We design a self-learning framework to extract several objects of interest simultaneously from Computed Tomography (CT) images. Our segmentation method has a learning phase that is based on reinforcement learning (RL) system. Each RL agent works on a particular sub-image of an input image to find a suitable value for each object in it. The RL system is define by state, action and reward. We defined some actions for each state in the sub-image. A reward function computes reward for each action of the RL agent. Finally, the valuable information, from discovering all states of the interest objects, will be stored in a Q-matrix and the final result can be applied in segmentation of similar images. The experimental results for cranial CT images demonstrated segmentation accuracy above 95%.     

Model-based reinforcement learning for partially observable games with sampling-based state estimation

Games constitute a challenging domain of reinforcement learning (RL) for acquiring strategies because many of them include multiple players and many unobservable variables in a large state space. The difficulty of solving such realistic multiagent problems with partial observability arises mainly from the fact that the computational cost for the estimation and prediction in the whole state space, including unobservable variables, is too heavy. To overcome this intractability and enable an agent to learn in an unknown environment, an effective approximation method is required with explicit learning of the environmental model. We present a model-based RL scheme for large-scale multiagent problems with partial observability and apply it to a card game, hearts. This game is a well-defined example of an imperfect information game and can be approximately formulated as a partially observable Markov decision process (POMDP) for a single learning agent. To reduce the computational cost, we use a sampling technique in which the heavy integration required for the estimation and prediction can be approximated by a plausible number of samples. Computer simulation results show that our method is effective in solving such a difficult, partially observable multiagent problem.     

Partially Observable Markov Decision Processes With Reward Information: Basic Ideas and Models

In a partially observable Markov decision process (POMDP), if the reward can be observed at each step, then the observed reward history contains information on the unknown state. This information, in addition to the information contained in the observation history, can be used to update the state probability distribution. The policy thus obtained is called a reward-information policy (RI-policy); an optimal RI-policy performs no worse than any normal optimal policy depending only on the observation history. The above observation leads to four different problem-formulations for POMDPs depending on whether the reward function is known and whether the reward at each step is observable. This exploratory work may attract attention to these interesting problems     

Learning policies for single machine job dispatching

Reinforcement learning (RL) has received some attention in recent years from agent-based researchers because it deals with the problem of how an autonomous agent can learn to select proper actions for achieving its goals through interacting with its environment. Each time after an agent performs an action, the environment's response, as indicated by its new state, is used by the agent to reward or penalize its action. The agent's goal is to maximize the total amount of reward it receives over the long run. Although there have been several successful examples demonstrating the usefulness of RL, its application to manufacturing systems has not been fully explored. In this study, a single machine agent employs the Q-learning algorithm to develop a decision-making policy on selecting the appropriate dispatching rule from among three given dispatching rules. The system objective is to minimize mean tardiness. This paper presents a factorial experiment design for studying the settings used to apply Q-learning to the single machine dispatching rule selection problem. The factors considered in this study include two related to the agent's policy table design and three for developing its reward function. This study not only investigates the main effects of this Q-learning application but also provides recommendations for factor settings and useful guidelines for future applications of Q-learning to agent-based production scheduling.     

Reinforcement Learning for feedback-enabled cyber resilience

The rapid growth in the number of devices and their connectivity has enlarged the attack surface and made cyber systems more vulnerable. As attackers become increasingly sophisticated and resourceful, mere reliance on traditional cyber protection, such as intrusion detection, firewalls, and encryption, is insufficient to secure the cyber systems. Cyber resilience provides a new security paradigm that complements inadequate protection with resilience mechanisms. A Cyber-Resilient Mechanism (CRM) adapts to the known or zero-day threats and uncertainties in real-time and strategically responds to them to maintain the critical functions of the cyber systems in the event of successful attacks. Feedback architectures play a pivotal role in enabling the online sensing, reasoning, and actuation process of the CRM. Reinforcement Learning (RL) is an important gathering of algorithms that epitomize the feedback architectures for cyber resilience. It allows the CRM to provide dynamic and sequential responses to attacks with limited or without prior knowledge of the environment and the attacker. In this work, we review the literature on RL for cyber resilience and discuss the cyber-resilient defenses against three major types of vulnerabilities, i.e., posture-related, information-related, and human-related vulnerabilities. We introduce moving target defense, defensive cyber deception, and assistive human security technologies as three application domains of CRMs to elaborate on their designs. The RL algorithms also have vulnerabilities themselves. We explain the major vulnerabilities of RL and present develop several attack models where the attacker target the information exchanged between the environment and the agent: the rewards, the state observations, and the action commands. We show that the attacker can trick the RL agent into learning a nefarious policy with minimum attacking effort. The paper introduces several defense methods to secure the RL-enabled systems from these attacks. However, there is still a lack of works that focuses on the defensive mechanisms for RL-enabled systems. Last but not least, we discuss the future challenges of RL for cyber security and resilience and emerging applications of RL-based CRMs.     

A Modified Memory-Based Reinforcement Learning Method for Solving POMDP Problems

Partially observable Markov decision processes (POMDP) provide a mathematical framework for agent planning under stochastic and partially observable environments. The classic Bayesian optimal solution can be obtained by transforming the problem into Markov decision process (MDP) using belief states. However, because the belief state space is continuous and multi-dimensional, the problem is highly intractable. Many practical heuristic based methods are proposed, but most of them require a complete POMDP model of the environment, which is not always practical. This article introduces a modified memory-based reinforcement learning algorithm called modified U-Tree that is capable of learning from raw sensor experiences with minimum prior knowledge. This article describes an enhancement of the original U-Tree’s state generation process to make the generated model more compact, and also proposes a modification of the statistical test for reward estimation, which allows the algorithm to be benchmarked against some traditional model-based algorithms with a set of well known POMDP problems.