Open AccessLEARNING FROM THE ENVIRONMENT WITH A UNIVERSAL REINFORCEMENT FUNCTION
Author(s) -
Diego Ariel Bendersky,
Juan E. Santos
Publication year - 2014
Publication title -
computing
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.184
H-Index - 11
eISSN - 2312-5381
pISSN - 1727-6209
DOI - 10.47839/ijc.5.3.410
Subject(s) - reinforcement learning , task (project management) , reinforcement , computer science , function (biology) , human–computer interaction , artificial intelligence , psychology , engineering , social psychology , evolutionary biology , biology , systems engineering
Traditionally, in Reinforcement Learning, the specification of the task is contained in the reinforcement function (RF), and each new task requires the definition of a new RF. But in the nature, explicit reward signals are limited, and the characteristics of the environment affects not only “how” animals perform particular tasks, but also “what” skills an animal will develop during its life. In this work, we propose a novel use of Reinforcement Learning that consists in the learning of different abilities or skills, based on the characteristics of the environment, using a fixed and universal reinforcement function. We also show a method to build a RF for a skill using information from the optimal policy learned in a particular environment and we prove that this method is correct, i.e., the RF constructed in this way produces the same optimal policy.