Rule Abstraction and Transfer in Reinforcement Learning by Decision Tree

被引：0

作者：

Wu, Min ^{[1
]}

Yamashita, Atsushi ^{[2
]}

Asama, Hajime ^{[1
]}

机构：

[1] Univ Tokyo, Dept Precis Engn, Hongo 7-3-1, Tokyo, Japan

[2] Univ Tokyo, Fac Precis Engn, Tokyo, Japan

来源：

2012 IEEE/SICE INTERNATIONAL SYMPOSIUM ON SYSTEM INTEGRATION (SII) | 2012年

关键词：

D O I：

暂无

中图分类号：

TP301 [理论、方法];

学科分类号：

081202 ;

摘要：

Reinforcement learning agents store their knowledge such as state-action value in look-up tables. However, loop-up table requires large memory space when number of states become large. Learning from look-up table is tabularasa therefore is very slow. To overcome this disadvantage, generalization methods are used to abstract knowledge. In this paper, decision tree technology is used to enable the agent to represent abstract knowledge in rule from during learning progress and form rule base for each individual task.

引用

页码：529 / 534

页数：6

共 13 条

[1] BEOM HR, 1995, IEEE T SYST MAN CYB, V25, P464, DOI 10.1109/21.364859
[2] Berenji H.R., 2002, IEEE T NEURAL NETWOR, V3, P724
[3] Bradtke S.J., 1995, NIPS 7
[4] Cassandra Anthony R., 1998, AAAI 1998 FALL S PLA, V1724
[5] Relational reinforcement learning
Dzeroski, S
De Raedt, L
Driessens, K
[J]. MACHINE LEARNING, 2001, 43 (1-2) : 7 - 52
[6] Faria G, 2000, NINTH IEEE INTERNATIONAL CONFERENCE ON FUZZY SYSTEMS (FUZZ-IEEE 2000), VOLS 1 AND 2, P847, DOI 10.1109/FUZZY.2000.839142
[7] Fuzzy decision trees: Issues and methods
Janikow, CZ
[J]. IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICS PART B-CYBERNETICS, 1998, 28 (01): : 1 - 14
[8] KAELBLING LP, 1996, MACH LEARN, V4, P237
[9] McCallum R., 1993, OV INC PERC UT DIST
[10] Pyeatt L. D., 2001, P 3 INT S AD SYST EV, V2, P7077

← 1 2 →