Simple and complex behavior learning using behavior hidden Markov model and CobART


Creative Commons License

Seyhan S. S., Alpaslan F. N., Yavaş M.

NEUROCOMPUTING, vol.103, pp.121-131, 2013 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 103
  • Publication Date: 2013
  • Doi Number: 10.1016/j.neucom.2012.09.013
  • Journal Name: NEUROCOMPUTING
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Page Numbers: pp.121-131
  • Keywords: Robot behavior learning, Hidden Markov model, Artificial neural network, CobART, IMITATION, CODES
  • Middle East Technical University Affiliated: Yes

Abstract

This paper proposes behavior learning and generation models for simple and complex behaviors of robots using unsupervised learning methods. While the simple behaviors are modeled by simple-behavior learning model (SBLM), complex behaviors are modeled by complex-behavior learning model (CBLM) which uses previously learned simple or complex behaviors. Both models include behavior categorization, behavior modeling, and behavior generation phases. In the behavior categorization phase, sensory data are categorized using correlation based adaptive resonance theory (CobART) network that generates motion primitives corresponding to robot's base abilities. In the behavior modeling phase, a modified version of hidden Markov model (HMM), is called Behavior-HMM, is used to model the relationships among the motion primitives in a finite state stochastic network. At the same time, a motion generator which is an artificial neural network (ANN) is trained for each motion primitive to learn essential robot motor commands. In the behavior generation phase, a motion primitive sequence that can perform the desired task is generated according to the previously learned Behavior-HMMs at the higher level. Then, in the lower level, these motion primitives are executed by the motion generator which is specifically trained for the corresponding motion primitive. The transitions between the motion primitives are done according to observed sensory data and probabilistic weights assigned to each transition during the learning phase. The proposed models are not constructed for one specific behavior, but are intended to be bases for all behaviors. The behavior learning capabilities of the model is extended by integrating previously learned behaviors hierarchically which is referred as CBLM. Hence, new behaviors can take advantage of already discovered behaviors. Performed experiments on a robot simulator show that simple and complex-behavior learning models can generate requested behaviors effectively. (C) 2012 Elsevier B.V. All rights reserved.