State abstraction in reinforcement learning

Bartosz Papis

Abstract

This work concerns state abstraction - one of commonly proposed solutions to the curse of dimensionality problem. A particular type of state abstraction - state space abstraction is analyzed as a variable selection issue. As an effect of this analysis, an incremental state abstraction algorithm is introduced, inspired by the notions of stimulus discrimination, ambiguity and closure from behavioral psychology. This algorithm correctly solves the variable selection problem by including or removing variables one by one. It is the first among existing solutions to work not only for discrete problems, but also continuous ones.
Rodzaj dyplomuPraca doktorska
Autor Bartosz Papis (WEiTI / IAiIS)
Bartosz Papis
- Instytut Automatyki i Informatyki Stosowanej
Tytuł w języku angielskimState abstraction in reinforcement learning
Języken angielski
Jednostka dyplomującaWydział Elektroniki i Technik Informacyjnych (WEiTI)
Dyscyplina naukiautomatyka i robotyka / dziedzina nauk technicznych / obszar nauk technicznych
Data obrony17-11-2015
Data zakończenia 24-11-2015
Promotor Andrzej Pacut (WEiTI / IAiIS)
Andrzej Pacut
- Instytut Automatyki i Informatyki Stosowanej

Recenzenci wewnętrzni Jarosław Arabas (WEiTI / ISE)
Jarosław Arabas
- Instytut Systemów Elektronicznych
Recenzenci zewnętrzni Jacek Koronacki
Jacek Koronacki
-
Wyróżnienietak
Paginacja 149
Słowa kluczowe w języku angielskimstate abstraction, Reinforcement Learning
Streszczenie w języku angielskimThis work concerns state abstraction - one of commonly proposed solutions to the curse of dimensionality problem. A particular type of state abstraction - state space abstraction is analyzed as a variable selection issue. As an effect of this analysis, an incremental state abstraction algorithm is introduced, inspired by the notions of stimulus discrimination, ambiguity and closure from behavioral psychology. This algorithm correctly solves the variable selection problem by including or removing variables one by one. It is the first among existing solutions to work not only for discrete problems, but also continuous ones.
Plik pracy
bpapis thesis.pdf 2.49 MB

Pobierz odnośnik do tego rekordu

Powrót