Shimon Whiteson's Adaptive Representations for Reinforcement Learning PDF

By Shimon Whiteson

ISBN-10: 3642139310

ISBN-13: 9783642139314

This ebook provides new algorithms for reinforcement studying, a sort of computer studying during which an self sustaining agent seeks a keep watch over coverage for a sequential selection activity. for the reason that present tools mostly depend upon manually designed answer representations, brokers that immediately adapt their very own representations have the aptitude to dramatically enhance functionality. This e-book introduces novel techniques for instantly studying high-performing representations. the 1st method synthesizes temporal distinction tools, the conventional method of reinforcement studying, with evolutionary equipment, which could examine representations for a extensive category of optimization difficulties. This synthesis is complete via customizing evolutionary easy methods to the online nature of reinforcement studying and utilizing them to adapt representations for worth functionality approximators. the second one method immediately learns representations according to piecewise-constant approximations of price features. It starts with coarse representations and steadily refines them in the course of studying, studying the present coverage and cost functionality to infer the easiest refinements. This e-book additionally introduces a singular process for devising enter representations. this system addresses the characteristic choice challenge by way of extending an set of rules that evolves the topology and weights of neural networks such that it evolves their inputs too. as well as introducing those new equipment, this publication offers large empirical leads to a number of domain names demonstrating that those suggestions can considerably enhance functionality over equipment with guide representations.

Show description

Read Online or Download Adaptive Representations for Reinforcement Learning PDF

Best nonfiction_6 books

Download e-book for iPad: Optimal control of distributed systems with conjugation by Ivan Vasilʹevich Sergienko; V S Deĭneka; Naum Zuselevich

"This paintings develops the method in accordance with which periods of discontinuous capabilities are utilized in order to enquire a correctness of boundary-value and preliminary boundary-value difficulties for the circumstances with elliptic, parabolic, pseudoparabolic, hyperbolic, and pseudohyperbolic equations and with elasticity concept equation platforms that experience nonsmooth ideas, together with discontinuous strategies.

Additional info for Adaptive Representations for Reinforcement Learning

Example text

When evolutionary methods are applied to reinforcement learning problems, they typically evolve a population of action selectors, each of which remains fixed during its fitness evaluation. The central insight behind evolutionary function approximation is that, if evolution is directed to evolve value functions instead, then those value functions can be updated, using TD methods, during each fitness evaluation. In this way, the system can evolve function approximators that are better able to learn via TD.

However, a handful of trials conducted for 200 generations verified that only very small additional improvements are made after 100 generations, without a qualitative effect on the results. Note that the progress of NEAT+Q consists of a series of 10,000-episode intervals. Each of these intervals corresponds to one generation and the changes within them are due to learning via Q-learning and backpropagation. Although each individual learns for only 100 episodes on average, NEAT’s system of randomly selecting individuals for evaluation causes that learning to be spread across the entire generation: each individual changes gradually during the generation as it is repeatedly evaluated.

In the server job scheduling domain, softmax NEAT+Q does perform better than softmax NEAT, though the difference is rather modest. Hence, in both domains, the most critical factor to boosting the performance of evolutionary computation is the use of an appropriate selection mechanism. 3 Comparing to Other Approaches The experiments presented thus far verify that the novel methods presented in this chapter can improve performance over the constituent techniques upon which they are built. This section presents experiments that compare the performance of the highest performing novel method, softmax NEAT+Q, to previous approaches.

Download PDF sample

Adaptive Representations for Reinforcement Learning by Shimon Whiteson


by Ronald
4.0

Rated 4.34 of 5 – based on 16 votes