Evolutionary algorithms and emergent intelligence

In order to perform adequately, knowledge-based artificial intelligence techniques rely on internal representations of the task environment. The requirement that this "explicit task knowledge" must be inside the agent leads to classic problems in AI: scaling, brittleness, learnability, knowledge acquisition, memory indexing and credit assignment. These problems are reduced or removed when the agent is allowed to interact with the task environment directly. In emergent intelligence, task specific knowledge emerges from the interaction of a simple agent and the original task environment. In effect, the task environment serves as a more efficient representation of the "explicit task knowledge," removing the need to represent it inside the agent. In this dissertation, evolutionary algorithms, computations that are modeled after natural selection, are analyzed and proposed to be a novel form of weak method that provide an ideal medium for implementing emergent intelligence. This dissertation also describes several experiments that demonstrate emergent intelligence during the acquisition of recurrent neural networks, finite state machines and modular LISP programs using a variety of evolutionary algorithms.