Tag Archives: Q-learning

On the effects of large variances in the transition function for Q-learning

D. Lee and W. B. Powell, Bias-Corrected Q-Learning With Multistate Extension. IEEE Transactions on Automatic Control, vol. 64, no. 10, pp. 4011-4023, DOI: 10.1109/TAC.2019.2912443.

Q-learning is a sample-based model-free algorithm that solves Markov decision problems asymptotically, but in finite time, it can perform poorly when random rewards and transitions result in large variance of value estimates. We pinpoint its cause to be the estimation bias due to the maximum operator in Q-learning algorithm, and present the evidence of max-operator bias in its Q value estimates. We then present an asymptotically optimal bias-correction strategy and construct an extension to bias-corrected Q-learning algorithm to multistate Markov decision processes, with asymptotic convergence properties as strong as those from Q-learning. We report the empirical performance of the bias-corrected Q-learning algorithm with multistate extension in two model problems: A multiarmed bandit version of Roulette and an electricity storage control simulation. The bias-corrected Q-learning algorithm with multistate extension is shown to control max-operator bias effectively, where the bias-resistance can be tuned predictably by adjusting a correction parameter.

Nice summary of reinforcement learning in control (Adaptive Dynamic Programming) and the use of Q-learning plus NN approximators for solving a control problem under a game theory framework

Kyriakos G. Vamvoudakis, Non-zero sum Nash Q-learning for unknown deterministic continuous-time linear systems, Automatica, Volume 61, November 2015, Pages 274-281, ISSN 0005-1098, DOI: 10.1016/j.automatica.2015.08.017.

This work proposes a novel Q-learning algorithm to solve the problem of non-zero sum Nash games of linear time invariant systems with N -players (control inputs) and centralized uncertain/unknown dynamics. We first formulate the Q-function of each player as a parametrization of the state and all other the control inputs or players. An integral reinforcement learning approach is used to develop a model-free structure of N -actors/ N -critics to estimate the parameters of the N -coupled Q-functions online while also guaranteeing closed-loop stability and convergence of the control policies to a Nash equilibrium. A 4th order, simulation example with five players is presented to show the efficacy of the proposed approach.

Multi-agent Q-learning applied to the defense against DDoS attacks with some provisions for scaling

Kleanthis Malialisa, Sam Devlina & Daniel Kudenkoa, Distributed reinforcement learning for adaptive and robust network intrusion response, Connection Science, Volume 27, Issue 3, 2015, DOI: 10.1080/09540091.2015.1031082.

Distributed denial of service (DDoS) attacks constitute a rapidly evolving threat in the current Internet. Multiagent Router Throttling is a novel approach to defend against DDoS attacks where multiple reinforcement learning agents are installed on a set of routers and learn to rate-limit or throttle traffic towards a victim server. The focus of this paper is on online learning and scalability. We propose an approach that incorporates task decomposition, team rewards and a form of reward shaping called difference rewards. One of the novel characteristics of the proposed system is that it provides a decentralised coordinated response to the DDoS problem, thus being resilient to DDoS attacks themselves. The proposed system learns remarkably fast, thus being suitable for online learning. Furthermore, its scalability is successfully demonstrated in experiments involving 1000 learning agents. We compare our approach against a baseline and a popular state-of-the-art throttling technique from the network security literature and show that the proposed approach is more effective, adaptive to sophisticated attack rate dynamics and robust to agent failures.

Finding the common utility of actions in several tasks learnt in the same domain in order to reduce the learning cost of reinforcement learning

Rosman, B.; Ramamoorthy, S., Action Priors for Learning Domain Invariances, Autonomous Mental Development, IEEE Transactions on , vol.7, no.2, pp.107,118, June 2015, DOI: 10.1109/TAMD.2015.2419715.

An agent tasked with solving a number of different decision making problems in similar environments has an opportunity to learn over a longer timescale than each individual task. Through examining solutions to different tasks, it can uncover behavioral invariances in the domain, by identifying actions to be prioritized in local contexts, invariant to task details. This information has the effect of greatly increasing the speed of solving new problems. We formalise this notion as action priors, defined as distributions over the action space, conditioned on environment state, and show how these can be learnt from a set of value functions. We apply action priors in the setting of reinforcement learning, to bias action selection during exploration. Aggressive use of action priors performs context based pruning of the available actions, thus reducing the complexity of lookahead during search. We additionally define action priors over observation features, rather than states, which provides further flexibility and generalizability, with the additional benefit of enabling feature selection. Action priors are demonstrated in experiments in a simulated factory environment and a large random graph domain, and show significant speed ups in learning new tasks. Furthermore, we argue that this mechanism is cognitively plausible, and is compatible with findings from cognitive psychology.

Reinforcement learning used for an adaptive attention mechanism, and integrated in an architecture with both top-down and bottom-up vision processing

Ognibene, D.; Baldassare, G., Ecological Active Vision: Four Bioinspired Principles to Integrate Bottom–Up and Adaptive Top–Down Attention Tested With a Simple Camera-Arm Robot, Autonomous Mental Development, IEEE Transactions on , vol.7, no.1, pp.3,25, March 2015. DOI: 10.1109/TAMD.2014.2341351.

Vision gives primates a wealth of information useful to manipulate the environment, but at the same time it can easily overwhelm their computational resources. Active vision is a key solution found by nature to solve this problem: a limited fovea actively displaced in space to collect only relevant information. Here we highlight that in ecological conditions this solution encounters four problems: 1) the agent needs to learn where to look based on its goals; 2) manipulation causes learning feedback in areas of space possibly outside the attention focus; 3) good visual actions are needed to guide manipulation actions, but only these can generate learning feedback; and 4) a limited fovea causes aliasing problems. We then propose a computational architecture (“BITPIC”) to overcome the four problems, integrating four bioinspired key ingredients: 1) reinforcement-learning fovea-based top-down attention; 2) a strong vision-manipulation coupling; 3) bottom-up periphery-based attention; and 4) a novel action-oriented memory. The system is tested with a simple simulated camera-arm robot solving a class of search-and-reach tasks involving color-blob “objects.” The results show that the architecture solves the problems, and hence the tasks, very efficiently, and highlight how the architecture principles can contribute to a full exploitation of the advantages of active vision in ecological conditions.

A new variant of Q-learning that alleviates its slow learning speed (with a brief review of reinforcement learning algorithms)

J.C. van Rooijen, I. Grondman, R. Babuška, Learning rate free reinforcement learning for real-time motion control using a value-gradient based policy, Mechatronics, Volume 24, Issue 8, December 2014, Pages 966-974, ISSN 0957-4158. DOI: 10.1016/j.mechatronics.2014.05.007

Reinforcement learning (RL) is a framework that enables a controller to find an optimal control policy for a task in an unknown environment. Although RL has been successfully used to solve optimal control problems, learning is generally slow. The main causes are the inefficient use of information collected during interaction with the system and the inability to use prior knowledge on the system or the control task. In addition, the learning speed heavily depends on the learning rate parameter, which is difficult to tune.
In this paper, we present a sample-efficient, learning-rate-free version of the Value-Gradient Based Policy (VGBP) algorithm. The main difference between VGBP and other frequently used algorithms, such as Sarsa, is that in VGBP the learning agent has a direct access to the reward function, rather than just the immediate reward values. Furthermore, the agent learns a process model. This enables the algorithm to select control actions by optimizing over the right-hand side of the Bellman equation. We demonstrate the fast learning convergence in simulations and experiments with the underactuated pendulum swing-up task. In addition, we present experimental results for a more complex 2-DOF robotic manipulator.