CPRMs: A Joint Model for Understanding and Action
Leon Rubin Barrett and Jerome A. Feldman
Operating in the real world requires both summarizing the world and taking actions. Summarizing the world has been addressed by statistical models like Bayes Nets and a first-order version called Probabilistic Relational Models (PRMs) . Unfortunately, these are not particularly appropriate for representing actions, which are compositional and partly deterministic. The Executable Schema (X-Schema), based on the well-known Petri Net, is a better representation of action . I intend to develop a principled way to combine a PRM, a perceptual interface with the world, with an X-Schema, a model of internal state and action. This combination is called a Coordinated PRM, or CPRM. I will not only formalize and analyze this structure, but also apply it to a real-world problem. For instance, embodied representations of actions are needed for language understanding, as it is hard to claim that a system understands the word "walk" if it lacks a model of the action. Furthermore, this combination would be useful in (real or simulated) robotics for the same reason combined action/world models are useful for humans. Finally, by exploring how humans represent goals and the world around them, this research could be a large step forward for cognitive science.
- D. Poole, "First-Order Probabilistic Inference," International Joint Conference on Artificial Intelligence, 2003, pp. 985-991.
- S. Narayanan, "Reasoning About Actions in Narrative Undertanding," Proc. Int. Joint Conf. Artificial Intelligence, Stockholm, Sweden, August 1999, pp. 350-358.