|
Accession number;01A0340297
|
| Title;Autonomous Composition of the State Space in Reinforcement Learning. |
| Author;
SHIBA TAKEMASA
(Nippon Inst. of Technol. Facul. of Eng.)
ISHIKAWA TAKASHI
(Nippon Inst. of Technol. Facul. of Eng.)
|
Journal Title;IEIC Technical Report (Institute of Electronics, Information and Communication Engineers)
|
Journal Code:S0532B
|
ISSN:0913-5685
|
|
VOL.100;NO.530(AI2000 56-65);PAGE.51-56(2001)
|
| Figure&Table&Reference;FIG.6, REF.6 |
| Pub. Country;Japan |
| Language;Japanese |
| Abstract;Reinforcement learning acquires the optimum action rules with the agent himself for the remuneration given from the environment. Since it is decided from the state which an agent can recognize, if the action rules for which it is asked has not conformed the optimum state space, it cannot gain the optimum action rules. There is a technique into which the agent itself divides state space autonomously as the method of solving this problem. This paper, describes a method of dividing state space based on the forecasting model of the state change using statistical alignment approximation as the technique of constituting state space autonomously. When environment changes, the improvement for building state space is described. Moreover, the improved technique is applied to the cart and pole problem and the result is shown. (author abst.) |
|
|
|
Related Articles;
|
