scholarly journals Entropic Regularization of Markov Decision Processes

Entropy ◽  
2019 ◽  
Vol 21 (7) ◽  
pp. 674
Author(s):  
Boris Belousov ◽  
Jan Peters
Keyword(s):  
Data Gathering ◽  
Direct Interaction ◽  
Learning Problems ◽  
Function Estimation ◽  
Policy Improvement ◽  
Reward Function ◽  
Learning Agent ◽  
Markov Decision ◽  
Entropic Regularization ◽  
Policy Optimization

An optimal feedback controller for a given Markov decision process (MDP) can in principle be synthesized by value or policy iteration. However, if the system dynamics and the reward function are unknown, a learning agent must discover an optimal controller via direct interaction with the environment. Such interactive data gathering commonly leads to divergence towards dangerous or uninformative regions of the state space unless additional regularization measures are taken. Prior works proposed bounding the information loss measured by the Kullback–Leibler (KL) divergence at every policy improvement step to eliminate instability in the learning dynamics. In this paper, we consider a broader family of f-divergences, and more concretely α -divergences, which inherit the beneficial property of providing the policy improvement step in closed form at the same time yielding a corresponding dual objective for policy evaluation. Such entropic proximal policy optimization view gives a unified perspective on compatible actor-critic architectures. In particular, common least-squares value function estimation coupled with advantage-weighted maximum likelihood policy improvement is shown to correspond to the Pearson χ 2 -divergence penalty. Other actor-critic pairs arise for various choices of the penalty-generating function f. On a concrete instantiation of our framework with the α -divergence, we carry out asymptotic analysis of the solutions for different values of α and demonstrate the effects of the divergence function choice on common standard reinforcement learning problems.

scholarly journals Object Affordance Driven Inverse Reinforcement Learning Through Conceptual Abstraction and Advice

2018 ◽  
Vol 9 (1) ◽  
pp. 277-294 ◽  
Author(s):  
Rupam Bhattacharyya ◽  
Shyamanta M. Hazarika
Keyword(s):  
Reinforcement Learning ◽  
High Dimensional ◽  
Inverse Reinforcement Learning ◽  
Intent Recognition ◽  
Reward Function ◽  
Object Affordances ◽  
Learning Agent ◽  
Markov Decision ◽  
Observed Behaviour ◽  
Object Affordance

Abstract Within human Intent Recognition (IR), a popular approach to learning from demonstration is Inverse Reinforcement Learning (IRL). IRL extracts an unknown reward function from samples of observed behaviour. Traditional IRL systems require large datasets to recover the underlying reward function. Object affordances have been used for IR. Existing literature on recognizing intents through object affordances fall short of utilizing its true potential. In this paper, we seek to develop an IRL system which drives human intent recognition along with the capability to handle high dimensional demonstrations exploiting the capability of object affordances. An architecture for recognizing human intent is presented which consists of an extended Maximum Likelihood Inverse Reinforcement Learning agent. Inclusion of Symbolic Conceptual Abstraction Engine (SCAE) along with an advisor allows the agent to work on Conceptually Abstracted Markov Decision Process. The agent recovers object affordance based reward function from high dimensional demonstrations. This function drives a Human Intent Recognizer through identification of probable intents. Performance of the resulting system on the standard CAD-120 dataset shows encouraging result.

Aero-engine acceleration control using deep reinforcement learning with phase-based reward function

2021 ◽  
pp. 095441002110462
Author(s):  
Qian-Kun Hu ◽  
Yong-Ping Zhao
Keyword(s):  
Reinforcement Learning ◽  
Trust Region ◽  
Engine Control ◽  
Control Task ◽  
Q Learning ◽  
Reward Function ◽  
Engine Control System ◽  
Aero Engine ◽  
Markov Decision ◽  
Policy Optimization

In this paper, the conventional aero-engine acceleration control task is formulated into a Markov Decision Process (MDP) problem. Then, a novel phase-based reward function is proposed to enhance the performance of deep reinforcement learning (DRL) in solving feedback control tasks. With that reward function, an aero-engine controller based on Trust Region Policy Optimization (TRPO) is developed to improve the aero-engine acceleration performance. Four comparison simulations were conducted to verify the effectiveness of the proposed methods. The simulation results show that the phase-based reward function helps to eliminate the oscillation problem of the aero-engine control system, which is caused by the traditional goal-based reward function when DRL is applied to the aero-engine control. And the TRPO controller outperforms deep Q-learning (DQN) and the proportional-integral-derivative (PID) in the aero-engine acceleration control task. Compared to DQN and PID controller, the acceleration time of aero-engine is decreased by 0.6 and 2.58 s, respectively, and the aero-engine acceleration performance is improved by 16.8 and 46.4 % each.

scholarly journals Reinforcement Learning with a Corrupted Reward Channel

2017 ◽  
Author(s):  
Tom Everitt ◽  
Victoria Krakovna ◽  
Laurent Orseau ◽  
Shane Legg
Keyword(s):  
Reinforcement Learning ◽  
Real World ◽  
Decision Problem ◽  
Inverse Reinforcement Learning ◽  
Markov Decision Problem ◽  
Software Bugs ◽  
Reward Function ◽  
Learning Agent ◽  
Markov Decision ◽  
Simplifying Assumptions

No real-world reward function is perfect. Sensory errors and software bugs may result in agents getting higher (or lower) rewards than they should. For example, a reinforcement learning agent may prefer states where a sensory error gives it the maximum reward, but where the true reward is actually small. We formalise this problem as a generalised Markov Decision Problem called Corrupt Reward MDP. Traditional RL methods fare poorly in CRMDPs, even under strong simplifying assumptions and when trying to compensate for the possibly corrupt rewards. Two ways around the problem are investigated. First, by giving the agent richer data, such as in inverse reinforcement learning and semi-supervised reinforcement learning, reward corruption stemming from systematic sensory errors may sometimes be completely managed. Second, by using randomisation to blunt the agent's optimisation, reward corruption can be partially managed under some assumptions.

scholarly journals Inverse reinforcement learning in contextual MDPs

2021 ◽  
Author(s):  
Stav Belogolovsky ◽  
Philip Korsunsky ◽  
Shie Mannor ◽  
Chen Tessler ◽  
Tom Zahavy
Keyword(s):  
Reinforcement Learning ◽  
Optimization Problem ◽  
Decision Processes ◽  
Inverse Reinforcement Learning ◽  
Convex Optimization Problem ◽  
Reward Function ◽  
Dynamic Treatment Regime ◽  
Markov Decision ◽  
Dynamic Treatment ◽  
Recorded Data

AbstractWe consider the task of Inverse Reinforcement Learning in Contextual Markov Decision Processes (MDPs). In this setting, contexts, which define the reward and transition kernel, are sampled from a distribution. In addition, although the reward is a function of the context, it is not provided to the agent. Instead, the agent observes demonstrations from an optimal policy. The goal is to learn the reward mapping, such that the agent will act optimally even when encountering previously unseen contexts, also known as zero-shot transfer. We formulate this problem as a non-differential convex optimization problem and propose a novel algorithm to compute its subgradients. Based on this scheme, we analyze several methods both theoretically, where we compare the sample complexity and scalability, and empirically. Most importantly, we show both theoretically and empirically that our algorithms perform zero-shot transfer (generalize to new and unseen contexts). Specifically, we present empirical experiments in a dynamic treatment regime, where the goal is to learn a reward function which explains the behavior of expert physicians based on recorded data of them treating patients diagnosed with sepsis.

PQROM: To optimize software defined network QoS-aware routing with proximal policy optimization

2021 ◽  
pp. 1-10
Author(s):  
Wei Zhou ◽  
Xing Jiang ◽  
Bingli Guo (Member, IEEE) ◽  
Lingyu Meng
Keyword(s):  
Software Defined Network ◽  
Training Time ◽  
Good Convergence ◽  
Reward Function ◽  
Routing Optimization ◽  
Policy Gradient ◽  
Discrete Action ◽  
Policy Optimization ◽  
Optimization Mechanism ◽  
Network Pattern

Currently, Quality-of-Service (QoS)-aware routing is one of the crucial challenges in Software Defined Network (SDN). The QoS performances, e.g. latency, packet loss ratio and throughput, must be optimized to improve the performance of network. Traditional static routing algorithms based on Open Shortest Path First (OSPF) could not adapt to traffic fluctuation, which may cause severe network congestion and service degradation. Central intelligence of SDN controller and recent breakthroughs of Deep Reinforcement Learning (DRL) pose a promising solution to tackle this challenge. Thus, we propose an on-policy DRL mechanism, namely the PPO-based (Proximal Policy Optimization) QoS-aware Routing Optimization Mechanism (PQROM), to achieve a general and re-customizable routing optimization. PQROM can dynamically update the routing calculation by adjusting the reward function according to different optimization objectives, and it is independent of any specific network pattern. Additionally, as a black-box one-step optimization, PQROM is qualified for both continuous and discrete action space with high-dimensional input and output. The OMNeT ++ simulation experiment results show that PQROM not only has good convergence, but also has better stability compared with OSPF, less training time and simpler hyper-parameters adjustment than Deep Deterministic Policy Gradient (DDPG) and less hardware consumption than Asynchronous Advantage Actor-Critic (A3C).

AUTOMATED VULNERABILITY SEARCH IN A WEB APPLICATION BASED ON REINFORCEMENT LEARNING

2021 ◽  
Vol 53 (1) ◽  
pp. 91-97
Author(s):  
OLGA N. VYBORNOVA ◽  
◽  
ALEKSANDER N. RYZHIKOV ◽  
Keyword(s):  
Reinforcement Learning ◽  
Web Application ◽  
Web Applications ◽  
Subject Area ◽  
Learning Technology ◽  
Web Application Security ◽  
Vulnerability Scanner ◽  
Learning Agent ◽  
Markov Decision ◽  
Python Programming

We analyzed the urgency of the task of creating a more efficient (compared to analogues) means of automated vulnerability search based on modern technologies. We have shown the similarity of the vulnerabilities identifying process with the Markov decision-making process and justified the feasibility of using reinforcement learning technology for solving this problem. Since the analysis of the web application security is currently the highest priority and in demand, within the framework of this work, the application of the mathematical apparatus of reinforcement learning with to this subject area is considered. The mathematical model is presented, the specifics of the training and testing processes for the problem of automated vulnerability search in web applications are described. Based on an analysis of the OWASP Testing Guide, an action space and a set of environment states are identified. The characteristics of the software implementation of the proposed model are described: Q-learning is implemented in the Python programming language; a neural network was created to implement the learning policy using the tensorflow library. We demonstrated the results of the Reinforcement Learning agent on a real web application, as well as their comparison with the report of the Acunetix Vulnerability Scanner. The findings indicate that the proposed solution is promising.

scholarly journals Research on the Multiagent Joint Proximal Policy Optimization Algorithm Controlling Cooperative Fixed-Wing UAV Obstacle Avoidance

Sensors ◽  
2020 ◽  
Vol 20 (16) ◽  
pp. 4546
Author(s):  
Weiwei Zhao ◽  
Hairong Chu ◽  
Xikui Miao ◽  
Lihong Guo ◽  
Honghai Shen ◽  
...  
Keyword(s):  
Reinforcement Learning ◽  
Attitude Control ◽  
Cooperative Control ◽  
Learning Algorithm ◽  
State Equations ◽  
Learning Agent ◽  
Environmental Adaptability ◽  
Decentralized Execution ◽  
Policy Optimization ◽  
Multi Uav

Multiple unmanned aerial vehicle (UAV) collaboration has great potential. To increase the intelligence and environmental adaptability of multi-UAV control, we study the application of deep reinforcement learning algorithms in the field of multi-UAV cooperative control. Aiming at the problem of a non-stationary environment caused by the change of learning agent strategy in reinforcement learning in a multi-agent environment, the paper presents an improved multiagent reinforcement learning algorithm—the multiagent joint proximal policy optimization (MAJPPO) algorithm with the centralized learning and decentralized execution. This algorithm uses the moving window averaging method to make each agent obtain a centralized state value function, so that the agents can achieve better collaboration. The improved algorithm enhances the collaboration and increases the sum of reward values obtained by the multiagent system. To evaluate the performance of the algorithm, we use the MAJPPO algorithm to complete the task of multi-UAV formation and the crossing of multiple-obstacle environments. To simplify the control complexity of the UAV, we use the six-degree of freedom and 12-state equations of the dynamics model of the UAV with an attitude control loop. The experimental results show that the MAJPPO algorithm has better performance and better environmental adaptability.

Hierarchical Reinforcement Learning

2011 ◽  
pp. 825-830
Author(s):  
Carlos Diuk ◽  
Michael Littman
Keyword(s):  
Reinforcement Learning ◽  
Learning Problems ◽  
Underlying Structure ◽  
Sequential Decision ◽  
State Spaces ◽  
Hierarchical Reinforcement Learning ◽  
Markov Decision ◽  
Finite Set ◽  
State Abstraction ◽  
Main Ideas

Reinforcement learning (RL) deals with the problem of an agent that has to learn how to behave to maximize its utility by its interactions with an environment (Sutton & Barto, 1998; Kaelbling, Littman & Moore, 1996). Reinforcement learning problems are usually formalized as Markov Decision Processes (MDP), which consist of a finite set of states and a finite number of possible actions that the agent can perform. At any given point in time, the agent is in a certain state and picks an action. It can then observe the new state this action leads to, and receives a reward signal. The goal of the agent is to maximize its long-term reward. In this standard formalization, no particular structure or relationship between states is assumed. However, learning in environments with extremely large state spaces is infeasible without some form of generalization. Exploiting the underlying structure of a problem can effect generalization and has long been recognized as an important aspect in representing sequential decision tasks (Boutilier et al., 1999). Hierarchical Reinforcement Learning is the subfield of RL that deals with the discovery and/or exploitation of this underlying structure. Two main ideas come into play in hierarchical RL. The first one is to break a task into a hierarchy of smaller subtasks, each of which can be learned faster and easier than the whole problem. Subtasks can also be performed multiple times in the course of achieving the larger task, reusing accumulated knowledge and skills. The second idea is to use state abstraction within subtasks: not every task needs to be concerned with every aspect of the state space, so some states can actually be abstracted away and treated as the same for the purpose of the given subtask.

EXISTENCE OF OPTIMAL STATIONARY POLICIES IN FINITE DYNAMIC PROGRAMS WITH NONNEGATIVE REWARDS

2001 ◽  
Vol 15 (4) ◽  
pp. 557-564 ◽  
Author(s):  
Rolando Cavazos-Cadena ◽  
Raúl Montes-de-Oca
Keyword(s):  
Control Policy ◽  
Stationary Policy ◽  
Reward Function ◽  
Total Reward ◽  
Dynamic Programs ◽  
Finite State ◽  
Markov Decision ◽  
Optimal Stationary Policy ◽  
Action Spaces ◽  
Discounted Criterion

This article concerns Markov decision chains with finite state and action spaces, and a control policy is graded via the expected total-reward criterion associated to a nonnegative reward function. Within this framework, a classical theorem guarantees the existence of an optimal stationary policy whenever the optimal value function is finite, a result that is obtained via a limit process using the discounted criterion. The objective of this article is to present an alternative approach, based entirely on the properties of the expected total-reward index, to establish such an existence result.

Sign in / Sign up

Export Citation Format

Share Document