Use this URL to cite or link to this record in EThOS:
Title: Online spatio-temporal learning and prediction for adaptive robotic systems : applications in tactile classification and learning assistance by demonstration
Author: Soh, Harold Soon Hong
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2013
Availability of Full Text:
Access from EThOS:
Access from Institution:
Successful biological systems adapt to change. Humans, for example, are capable of continual self-improvement and gain new skills with experience. Similar online learning characteristics would enable robotic systems to autonomously improve their capabilities over time. In this thesis, we focus on the problem of iteratively learning from multivariate time-series; the "raw material" that we use to make inferences about the future. We adopt a combined approach: gaining inspiration from biological systems, in particular recurrent neural networks, and merging these ideas with recent advances in statistical machine learning. The resulting algorithm --- the online echo-state Gaussian process (OESGP) --- learns in an online manner, produces predictive distributions and attains state-of-the-art results on a variety of benchmark problems. We further extend this method to networks of "infinite size" through a recursive kernel with automatic relevance determination. This allows for online optimisation of the hyper-parameters through stochastic natural gradient descent, which improves adaptability and alleviates the problem of reservoir parameter specification. Using this online infinite ESGP (OIESGP) as a building block, we address two challenging problems in robotics: online tactile learning using the iCub humanoid platform and smart mobility assistance on the ARTY smart wheelchair. For the former, we develop online generative and discriminative classifiers that learn new objects "on-the-fly" and refine older models with new sensory input. For the latter, we adopt a novel approach by applying imitation learning to derive assistive policies. We present an OIESGP-based probabilistic mixture model for learning when and how to appropriately assist, and demonstrate its effectiveness in simulation and real-world experiments with human subjects.
Supervisor: Demiris, Yiannis Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available