Evolutionary Algorithms for Intelligent Software Design

摘要

In this work we consider the notion of "intelligence" in its most general meaning, namely, as computational aspect of the ability to achieve goals. We present the result of our work on emergent intelligence of automatically developed software, viewed as ability of such software (when controlling realistically simulated artifact - Snakebot) to achieve goals. The emergent primitive intelligence is observed at the level of algorithmic paradigm (genetic programming) employed to design the controlling software of the-Snakebot. which achieve the goals without the need to be explicitly taught about how to do so. As an instant of primitive intelligence we consider the ability of snakebot to obtain fast (sidewinding) locomotion gaits. Moreover, at the mechanical level of the Snakebot, viewed as a complex mechanical system comprising many simple phenotypic segments, we discovered that such primitive intelligence (locomotion gaits) of the whole artifact emerge from relatively simply defined motion patters of the phenotypic segments. Empirical results verify the robustness of sidewinding Snakebot, viewed as the ability to retain its velocity when situated in unanticipated environment. Snakebot easily overcomes various types of obstacles such as a pile of or burial under boxes, rugged terrain and small walls. The adaptation of Snakebot to partial damage by gradually improving its velocity characteristics is discussed. Discovering compensatory locomotion traits, Snakebot recovers completely from single damage and recovers a major extent of its original velocity when more significant damage is inflicted.