Praktická aplikace Q-učení

Continuous Q-learning application

článek ve sborníku ve WoS nebo Scopus

en

Standard algorithm of Q-Learning is limited by discrete states and actions and Q-function is usually represented as discrete table. To avoid this obstacle and extend the use of Q-learning for continuous states and actions the algorithm must be modified and such modification is presented in the paper. Straightforward way is to replace discrete table with suitable approximator. A number of approximators can be used, with respect to memory and computational requirements the local approximator is particularly favorable. We have used Locally Weighted Regression (LWR) algorithm. The paper discusses advantages and disadvantages of modified algorithm demonstrated on simple control task.

Standardní algoritmus metody Q-učení je omezen používáním diskrétních stavů a akcí. V tomto případě je Q-funkce representována jako diskrétní tabulka. Metoda popisovaná v tomto příspěvku se snaží obejít problém s diskretizací tím, že je od počátku navržena jako spojitá. Diskrétní tabulka Q-hodnot je nahrazena vhodným aproximátorem. V tomto příspěvku hodnotíme výhody a nevýhody spojitého Q-učení oproti jeho diskrétní variantě.

Standard algorithm of Q-Learning is limited by discrete states and actions and Q-function is usually represented as discrete table. To avoid this obstacle and extend the use of Q-learning for continuous states and actions the algorithm must be modified and such modification is presented in the paper. Straightforward way is to replace discrete table with suitable approximator. A number of approximators can be used, with respect to memory and computational requirements the local approximator is particularly favorable. We have used Locally Weighted Regression (LWR) algorithm. The paper discusses advantages and disadvantages of modified algorithm demonstrated on simple control task.

Q-learning, Machine learning, Locall approximators

2004

10.05.2004

Institute of Thermonechanics Academy of Sciences of the Czech Republic, Prague 2004

Prague

80-85918-88-9

Engineering Mechanics 2004

1

2