07d7a2a293b9c704d2d411c1fed53ded.ppt

- Количество слайдов: 10

UW Contributions: Past and Future Martin V. Butz Department of Cognitive Psychology University of Würzburg, Germany http: //www-illigal. ge. uiuc. edu/~butz [email protected] uni-wuerburg. de 01/18/2005

Overview 1. 2. Publications Dissemination work 1. 2. 3. Work on multiple facets 1. 2. 3. 4. Continued work on the XCS classifier system for function approximation – hyperellipsoidal conditions, RLS, and compaction Visuomotor grounded tracking system The SURE_REACH architecture: A Sensorimotor Unsupervised Redundancy Resolving control Architecture Collaborations 1. 2. 3. 4. 06/15/2005 Organization of ABi. ALS 2006 during SAB with ISTC-CNR Workshop proceedings CD, Postworkshop book with additional overview papers OFAI with XCS / AIS comparisons ISTC-CNR with ideomotor principle & TOTE - related system comparisons Tracking and object-recognition experiments with IDSIA (+ LUND ? ) Integration of SURE_REACH architecture with RL component from ISTC-CNR Towards Hierarchical Cognitive Systems 2

Publications Book: Butz, M. V. (2006) Rule-based evolutionary online learning systems: A principled approach to LCS analysis and design. Studies in Fuzziness and Soft Computing Series, Springer Verlag, Berlin-Heidelberg, Germany. Journals: Butz, M. V. , Goldberg, D. E. , Lanzi, P. L. , & Sastry, K. (in press). Problem Solution Sustenance in XCS: Markov Chain Analysis of Niche Support Distributions and Consequent Computational Complexity. Genetic Programming and Evolvable Machines. Butz, M. V. , Pelikan, M. , Llorà, X. , & Goldberg, D. E. (in press) Automated global structure extraction for effective local building block processing in XCS. Evolutionary Computation Journal. Butz, M. V. , Herbort, O. , & Hoffmann, J. (submitted) Exploiting Redundancy for Flexible Behavior: Unsupervised Learning of a Modular Sensorimotor Control Architecture Butz, M. V. , Lanzi, P. L. , & Wilson, S. W. (submitted) Function Approximation with XCS: Hyperellipsoidal Conditions, Recursive Least Squares, and Compaction. IEEE Transactions on Systems Man and Cybernetics, Part B. Conferences: Butz, M. V. , Pelikan, M. (2006). Studying XCS/BOA learning in Boolean functions: Structure encoding and random boolean functions. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2006). 1449 -1456. Butz, M. V. , Lanzi, P. L. , Wilson, S. W. (2006). Hyper-ellipsoidal conditions in XCS: Rotation, linear approximation, and solution structure. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2006). 1457 -1464. Workshops: Pezzulo, G. , Baldassarre, G. , Butz, M. V. , Castelfranchi, C. , & Hoffmann, J. (2006) An Analysis of the Ideomotor Principle and TOTE. ABi. ALS 2006. Herbort, O. , & Butz, M. V. (2006). Unsupervised Learning of Inverse Dynamics Model. Cog. Sys II. 06/15/2005 Towards Hierarchical Cognitive Systems 3

Dissemination • … apart from publications: • Multiple invited presentations on XCS and Anticipations • Organization of ABi. ALS 2006 with ISTC-CNR: Anticipatory Behavior in Adaptive Learning Systems • Workshop CD • Post-workshop proceedings book Submission deadline, 30 th of November 06/15/2005 Towards Hierarchical Cognitive Systems 4

Work on XCSF for Function Approximation • XCSF is a partially overlapping, piecewise linear function approximation systems – Learns iteratively, online – Clusters the space to yield accurate function approximations • Improvements: – Rotation improves / speeds-up evolutionary process – RLS makes approximations more accurate – Compaction for generalized function approximation representation. • Results: – Effective function approximation – max problem is 7 D (sin(2 PI (x 1+…+x 7)) – Noise robustness – N(0, . 1) noise – best performance compared to results in ICML 2004 – Generalization: Comparable results to Atkeson, Schaal (1998) Neural Computation with heuristic approach. 06/15/2005 Towards Hierarchical Cognitive Systems 5

Visuomotor-grounded Tracking System Goal: Object recognition by analysis of observed visual (object-related) flow. Currently integrated in IKAROS Approach: 1. Learn about optical flow observing visual changes caused by own actions (retinal tracking movement) 1. Represent the flow by local, sigma-pi predictors 2. Integrate population encoding of motor action, current perception to generate future perception 2. Observe flow deriving local movement vectors 3. Use the movement information to differentiate types of objects 4. Use the knowledge to predict behavior of object 1. Bouncing behavior 2. Object permanence 06/15/2005 Towards Hierarchical Cognitive Systems 6

The SURE_REACH Architecture Anticipatory, goal-oriented robot arm approach Covers two spatial representations with population encoding: End-point coordinate space Posture space Associates End-point and posture space. Learns associative, action-dependent inverse models Reaches goals flexibly and accurately: Choosing the shortest path if possible. Flexibly obeying additional goal constraints. Avoiding obstacles by an effective inhibition of neurons 06/15/2005 Towards Hierarchical Cognitive Systems 7

Collaborations 1. OFAI with XCS / AIS comparisons – test in robot arm scenario 2. ISTC-CNR with ideomotor principle & TOTE related system comparisons 3. Tracking and object-recognition experiments with IDSIA (+ LUND ? ) 4. Integration of SURE_REACH architecture with RL component from ISTC-CNR (? ) 06/15/2005 Towards Hierarchical Cognitive Systems 8

Deliverable 4. 2 • Due date: Month 30 – that is, 03/2007 • Description: Results of tests of architectures which model mechanisms based on analogy, proactive and goal directed behaviour. The deliverable will be a report that illustrates the results of the tests of different architectures and of at least one robot and a comparison of their performances in the implemented scenario selected for analysing the cognitive functions set of Goal directed behaviour, Pro-activity and Analogy. Also a description of the implemented architectures and of the robot will be given. • Milestone 4. 2. Final results concerning measures of performance of improved architectures and comparison of performances. 06/15/2005 We will provide final results concerning measures of performance of at least two improved architectures in the scenario selected for analyzing the cognitive functions set of Goal directed behavior, Pro-activity and Analogy. At least one of the improved architectures will be a real Robot, the other ones will be simulated using Hierarchical Cognitive Systems Towards agent software. Moreover, a 9

For Deliverable 4. 2 • Who is doing system comparisons? • On which scenario can we test the systems for comparison? • Which real robot are we going to use for comparison? 06/15/2005 Towards Hierarchical Cognitive Systems 10