[1Q5-GS-11] Robot and real worlds: Fundamental theory

[1Q5-GS-11-03] Fast Sequence Pattern Extraction Using Slice Sampling

〇Masatoshi Nagano¹, Tomoaki Nakamura¹, Takayuki Nagai², Daichi Mochihashi³, Ichiro Kobayashi⁴, Wataru Takano² (1. The University of Electro-Communications, 2. Osaka University, 3. Institute of Statistical Mathematics, 4. Ochanomizu University)

Keywords:Segmentation, Unsupervised Learning, Gaussian Process-Hidden Semi-Markov Model, Slice Sampling, Forward filtering-Backward sampling

In statistical time-series modeling, the parameters can be estimated by using the Forward filtering-Backward sampling (FFBS). However, FFBS requires to compute the probabilities for all possible combinations of latent variables. Therefore, the computation of FFBS takes a high cost. To overcome this problem, we introduce Slice sampling (SS) into FFBS to stochastically restrict the combinations of latent variables and we propose Gaussian Process-Hidden Semi-Markov Model-STAR (GP-HSMM*), in which SS is introduced into GP-HSMM. GP-HSMM is a method for dividing continuous time-series data into significant segments and classifying them into classes based on similarities in an unsupervised manner. In the FFBS of GP-HSMM*, the probabilities of valid combinations of the latent variables are computed and the latent variables are sampled based on the computed probabilities. Next, the combinations of latent variables are restricted by using SS based on the sampled latent variables. By iterating this procedure, the proposed method can reduce a computational cost by computing probabilities of important combinations of latent variables for segmentation. In the experiment, we use synthetic and motion-capture data and show that our proposed method can efficiently estimate segments compared with GP-HSMM.