![]() |
Mark Witkowski's Home Page |
Keywords: Cognitive Robotics, Dynamic Expectancy Theory, Anticipatory Learning, Simulation of Adaptive Behaviour, Biologically Inspired Robotics, Action Selection Research Interests I am currently associated with Prof. Erol Gelenbe's Intelligent Systems and Networks group and have been working on understanding intentionality in a tactical environment. You can find a list of selected publications here. I have recently completed work in Murray Shanahan's Cognitive Robotics group. Our most recent, completed, EPSRC project "Abductive Robot Perception: Modelling Granularity and Attention in Euclidean Representational Space" takes an abductive (reasoning from observations to postulated causes) reasoning approach to active perception for robotics. Key results from this work include a formal view of the perception process and a consolidation of multiple aspects of attention and awareness in this formalised view. Perception is taken as a combination of bottom-up sense driven and top-down hypothesis driven activities. Features extracted from the robot sensors indicate explanatory hypotheses in the bottom-up phase. Subsequently, the deductive consequences of each of (the many, possibly competing) hypotheses are used to disambiguate between alternative interpretations of the sense data to provide an explanation of that data. This project uses our upper torso humanoid robot Ludwig. Our previous projects in the Imperial College Cognitive Robotics series have been Spatial Reasoning and Perception in a Humanoid Robot, Cognitive Robotics II and Cognitive Robotics I, which investigated the use of combines Shanahan's existing Event Calculus formalism in an abductive framework to investigate a range of topics in robot sensor data assimilation, planning and action selection. My on-going research interests include the simulation and understanding
of natural Behaviour
through anticipatory learning methods and associated action selection
strategies. My latest paper on these issues "An Action Selection
Calculus" is in the Adaptive
Behavior Journal. These are a continuation of earlier
work on the "Dynamic
Expectancy Theory", which integrates innate reactive (pre-programmed or evolved)
behaviours with goal-directed learned behaviour to address the action
selection problem. Learning in the Dynamic Expectancy Model is accomplished by an ingenious variant of the
reinforcement learning method in which expectancies about the environment
("µ-hypotheses") are created and corroborated (by "µ-experiment") by
a predictive method which dispenses with the need for the provision of external
reward to drive learning. Motivation and learning are separated and
learning may take place latently, independently of any known or assumed
goal. The system generates reactive goal-directed behaviours by formulating a
"dynamic policy
map" from stored µ-hypotheses when goals arise. I was local organisation chair for Towards Autonomous Robotics Conference (TAROS-05) at Imperial College in September 2005 (on-line archive), chair of the programme committee for TAROS-06 (proceedings, large file) held at Surrey University in September 2006, and conference co-chair for TAROS-07 held in Aberystwyth in September 2007 and TAROS-08 in Edinburgh, 1-3 September 2008. TAROS-09 will be in Ulster 31 August to 2 September 2009. Please contact me by e-mail at the Department or at AIQ Limited.
Thesis, full text in pdf and postscript
Some Photos of Past Robot Projects
|