Self-Segmentation of Sequences Algorithm with Eligibility Traces in POMDPs  

Oleh:

Kamaya, Hiroyuki ; Abe, Kenichi

Jenis: Article from Article
Dalam koleksi: Final Program and Book of Abstracts: The 4th Asian Control Conference, September 25-27, 2002 (Sep. 2002), page 408-413.
Topik: Self-Segmentation; Sequences Algorithm; Eligibility Traces; POMDPs; Partially Observable Markov Decision Processes
Fulltext: AC021697.PDF (181.81KB)

Isi artikel

Partially Observable Markov Decision Processes (POMDPs) provide a general decision-making framework for acting optimally in partially observable domains. Despite the powerful ability of POMDPs, their use is significantly limited due to the huge computational cost for an agent to find an optimal policy. Consequently, recent efforts have focused on the development of efficient algorithms to generate nearoptimal policies. Current algorithms, however, are still generally unable to handle slightly larger and noisier problems. We propose a new hierarchical reinforcement learning (RL) algorithm which uses eligibility traces to speed up learning, to make it more robust to hidden states, and to handle deterministic or nondeterministic problems. This on-line RL algorithm is called Self-Segmentation of Sequences algorithm with eligibility traces (SSS-ET), which extended Sun and Sessions’s original SSS algorithm. SSS-ET is compared with the original SSS in a partially observable navigation task. Finally, it is confirmed that the proposed SSS-ET is clearly outperforming the original SSS.

Opini Anda

Klik untuk menuliskan opini Anda tentang koleksi ini!

Kembali