A connectionist model for learning continuous temporal relationships in realtime
John Dale Morrison · 1992
Limitations of the current symbolic paradigm restrict its capacity to support manageable and verifiable knowledge base development for expert system simulations being applied to dynamic domains, thereby undermining the original motivations for rule-based simulations. Furthermore, these limitations restrain interest in the use of simulations to: assess behavioral influences on complex system performance; generate new rules of behavior for developmental systems; and develop, verify, and update intelligent system software. This dissertation argues that because expertise acquired in complex, spatio-temporal environments is associated with unique aspects of memory and action-response sequences that are resistant to lexical cues and expression, an alternative model is required for environments that are not structurally well understood, predictable, deterministic, symbolic, or temporally static. Motivated by an emerging body of research into connectionist, closed-loop, reinforcement learning models, this dissertation proposes and evaluates an alternative model within a complex, combat simulation. The unique characteristics of this prototype are that its memory units are truly autonomous (contain all system characteristics locally), that reinforcement signals are indistinguishable from sensory input, and that unbounded weight growth is controlled by a uniquely adaptive threshold filter. These characteristics support a general model of intelligence that is: naturally extensible; represents complex goal functions; and supports rule-based (deductive) reasoning in the presence of environmental activity that is consistent with expectation, as well as goal-based (inductive) reasoning in the presence of uncertainty--unfamiliar patterns of activity. The experiment demonstrates that the prototype is not only capable of supporting effective strategy refinement, but converging to stable, rule-based behavior quickly and efficiently without unbounded weight growth. These results motivate further research into development of an even more powerful model of intelligence that would support the application of intelligent simulations to the broader, long-term goals associated with developing knowledge-bases for developmental hardware and software systems.