There will be two seminars, 30 minutes each. Details follow:
TOPIC 1: TClass: A supervised, comprehensible concept learner for multivariate time series
SPEAKER 1: Mohammed Waleed Kadous, School of Computer Science and Engineering, University of New South Wales
DATE: Friday, 22nd June 2001 TIME: 12:00 - 13:00
PLACE: CSE K17 Basement CAT room 2 (B02)
TClass is a learner specifically designed for application to supervised temporal domains. As such, it eschews the more traditional attribute-value format. Rather, each instance is represented as a multivariate time series. The objective is to learn to classify different types of time series and to produce comprehensible descriptions of learnt concepts. Typical (and tested) application domains include: sign language recogniton, ECG classification and isolated word recognition. Its key novel contributions are that it produces human-readable definitions, high levels of accuracy and good data efficiency (learnt concepts require fewer training examples than traditional techniques to produce high levels of accuracy). Current results are extremely promising. In the sign language domain with 95 sample signs from Auslan, 22 channels of data with a duration of 100 to 200 samples and 23 samples per sign, we are obtaining accuracy in excess of 97 per cent. On a seven-class ECG problem, we are obtaining 68 per cent accuracy, very close to the performance of human experts with 71 per cent accuracy. We perform extremely well on artificial benchmarks suggested by other researchers. Whatês more, we do so while producing descriptions of the learnt concepts.
BIOGRAPHY OF SPEAKER 1:
Mohammed Waleed Kadous (Waleed) received his Bachelor of Engineering (Computer Engineering) and University Medal from the University of New South Wales in 1995. He is currently completing his PhD in the field of machine learning at UNSW, where he is an associate lecturer in Computer Graphics. Waleed has received several scholarships and awards for his work, including an Institution of Engineers of Australia for his work with sign language recognition and scholarships to attend AAAI and ICML2001. He is a part-time consultant on Linux, Perl and Java.
TOPIC 2: Learning Hierarchical Decomposition for Factored MDPs
SPEAKER 2: Bernhard Hengst , School of Computer Science and Engineering, University of New South Wales
–The biggest open problem in reinforcement learning is to discover hierarchical structure” (Dietterich , 2000). This talk describes an algorithm that automatically finds state abstractions and generates a hierarchical structure to represent and solve MDPs. The operation of the algorithm is demonstrated using a simple room example and Dietterichês taxi with fuel domain.
BIOGRAPHY OF SPEAKER 2:
Bernhard is a PhD student with research interests in machine learning and autonomous agents. For the last 25 years he has worked in private industry, initially in operations research with ICI Australia and ICI (UK) and later as general manager with ACI Australia, Ferntree Computer Corporation and GE Capital (CT, USA). Bernhard holds a B.Sc and B.E. with first class honours from the University of New South Wales.
School of Computer Science & Engineering, UNSW.