ICML Workshop on Predictive Representations of World Knowledge

[Rich Sutton]

                              Announcing
                         an ICML workshop on

              PREDICTIVE REPRESENTATIONS OF WORLD KNOWLEDGE

        Call for participation/submissions (http://prwk.rlai.net)

It has long been postulated that knowledge of the world ultimately 
comes down to predictions about what we will sense as a function of 
what we do. For example, in this view we know what is inside a box if 
we can predict what we would see if we were to open it. We know the 
three-dimensional shape of an object if we can predict its new 
appearance for each of various way in which we might rotate it. We 
might even be said to know that "Tweety is a bird" if we can make 
appropriate predictions about what we might see and hear (e.g., about 
feathers, flying, and chirps) if we were to meet Tweety. Such 
predictive representations of world knowledge have important potential 
advantages. If predictions are defined in terms of primitive sensations 
and actions (i.e., are grounded), then they can be directly compared 
with what actually does happen, enabling the knowledge to be verified 
or perhaps even learned without human intervention. If the predictions 
are deterministic, or Markov in an appropriate sense, then the 
knowledge can be immediately used in a variety of state-space planning 
methods.

For these and other reasons, understanding world knowledge in 
predictive, sensori-motor terms has been a long-standing goal of 
philosophy, psychology, and artificial intelligence. So far it has 
remained a distant goal, but recent progress in machine learning seems 
to bring nearer the possibility of addressing it productively with 
mathematics and computational studies. In particular, we are thinking 
of research on:

    Predictive representations of state (PSRs)
    Observable operator models (OOMs)
    Temporal abstraction in reinforcement learning (options, HAMs, MAXQ)
    Diversity-based induction of finite-state automata
    Deitic, or indexical, representations
    Signals to symbols, symbol grounding

The goal of this workshop is to bring together scientists interested in 
these and other topics to present, discuss, and make further progress 
toward understanding how knowledge of the world can be represented in 
predictive, sensori-motor terms.

Program Committee: Rich Sutton (University of Alberta), Satinder Singh 
(University of Michigan), Herbert Jaeger (International University 
Bremen, Germany), Michael Littman (Rutgers University), Peter Stone 
(University of Texas at Austin), Tim Oates (University of Maryland 
Baltimore County), Martin Butz (University of Illinois, not confirmed)

For details on submissions, due March 26, see http://prwk.rlai.net