Title: "The Dynamics Based Control Framework"
Sub-title: How to use the fact that we see only how we want to see

Speaker: Zinovi Rabinovich, Computer Science, Hebrew University
Date: Thursday, 19 April 2007
Time: 11am
Place: Ross 201

Dynamics Based Control (DBC) is a novel framework for continual
planning and control of stochastic environments. While it can be
related to the principles of model following and perceptual control,
DBC directly targets system dynamics, rather than the system
state. The DBC structure takes into account not only the computational
limitations of an agent, but directly employs such knowledge to
achieve its task. Being a general and flexible framework, DBC can
potentially have many algorithmic solutions. One of them, an
approximate greedy solution called Extended Markov Tracking (EMT), has
been shown to be operational in various tasks. Multi-target and
multi-agent variations of EMT, together with EMT's relative
computational efficiency, allow the algorithm, and its generalized DBC
framework, to stand alongside other popular alternatives as an
effective control methodology.

The talk will start with an analysis of several scenarios, which will
assist in building the intuition necessary for Dynamics Based
Control. We will then proceed to discuss a formalization of the DBC
framework, and its relationship to classical control and planning
approaches. Next, the EMT algorithm will be presented for discrete
Markovian environments, as well as the algorithm's extensions to
multi-agent and multi-target tasks, and their experimental support. I
will conclude with some theoretical and computational complexity
remarks.