Exploitation of environmentally-derived sensorimotor isomorphisms
Răzvan V. Florian
This contribution presents a biologically-inspired heuristic which situated, embodied artificial agents may use for solving advanced tasks, clasically considered to involve higher level reasoning. Through active sensorimotor interaction with their environment, the artificial agents adapt their cognitive structures to the structure of this environment, and internalize part of this structure. The internalized structure, based on neural networks, needs not to be explicit or representational. It reflects sensorimotor invariants encountered in past explorations of the environment, that dynamically influence future sensorimotor interactions.
Like in some current robotic implementations, the internalized structure can generate adaptive behavior that ranges from reactive responses to planned action. Simulation (forward modeling), on the basis of internalized structure, may allow a support for making choices between different actions, as a function of the agent's goals.
Further, this rich internalized structure can also be used, possibly adapted, for the emergence of abstract concepts, and for forward modeling and problem solving in domains where dense sensorimotor interaction or feedback are not easily available. The basis for this possibility of reuse is the possibility of partial isomorphisms between the structure of different domains. The validity of this heuristics is proved, among others, by the success of human science. For example, the concept of number emerges in humans because of partial isomorphisms of several sensorimotor domains (manipulation and construction of objects, measurement of linear dimensions and quantities, movement on a path (Lakoff and Nunez, 2000)). This is possible because there is a nontrivial physical structure of the world, that exhibits these isomorphisms; and because our sensorimotor apparatus is able to detect this structure (Florian, 2002). Artificial intelligent agents should also exploit these isomorphisms that are lingering in the environment. This heuristic may lead to advanced cognitive capabilities, similar to human analogical problem solving or scientific skills.
This approach has several advantages over classical AI approaches for solving abstract problems. Acquiring structure from sensorimotor interaction is a mechanism suitable to the distributed representation of knowledge, with neural networks, having known advantages over symbolic representation (graceful degradation, robustness, etc.). Complex structure is easier to acquire from the environment, rather than being preprogrammed by the designer of the artificial agent or evolved. To be preprogrammed, it may require unfeasible amounts of work for the programmer, and it would be biased by the ontology of the programmer. To be evolved, it may imply the exploration of a much too large search space. The emergence of structure from sensorimotor interaction is dependent on sensorimotor capabilities and environments; sensorimotor capabilities different of the human ones can lead to novel ontologies, and thus to novel solutions to certain problems (Florian, 2002). Unlike current approaches to analogical reasoning, based on search, this method, based on distributed representation, would allow dynamic restructuring of knowledge, and thus be a premise for the emergence of creativity in artificial cognitive agents.
Preliminary results of experimental work directed to implement this heuristic in animats will be presented.