Causally-Guided Evolutionary Computation for Design

Thumbnail Image


Publication or External Link






During recent years, evolutionary computation methods have been used successfully to discover solutions to problems involving design and invention in a wide variety of fields. However, for the evolutionary process to remain computationally tractable when applied to increasingly complex design problems, new extensions must be developed that increase the efficiency and effectiveness with which evolutionary systems produce optimal designs. To this end, the goal of the research presented here is to develop one such potential extension: causally-guided evolution. By this I mean evolutionary systems where the application of genetic operators to an individual are driven in part by observing that individual's performance characteristics and applying these operators based on explicit cause-effect relations in the domain. This differs from past evolutionary methods in which, after fitness-based selection, genetic operators are applied to individuals blindly and randomly (i.e., without respect to the performance characteristics of the individuals).

In this context, this dissertation makes a number of significant contributions. A framework for causally-guided evolution is defined, including causally-guided genetic operators based on causal knowledge that is supplied by domain experts. The ability of these methods and causally-guided mutation to produce better solutions than conventional evolutionary processes is demonstrated on a neural network optimization task. These methods are then extended to include crossover, and the synergistic effects of causally-guided crossover and mutation are demonstrated when applied to a real-world antenna design task. Causally-guided mutation is extended further to influence both where and how mutation occurs, and the effectiveness of this approach is shown when applied to a constructive design task that creates synthetic social networks. Finally, a causally-guided evolutionary system that acquires causal knowledge through observation of the evolutionary process, rather than being given the knowledge a priori, is developed and successfully applied, demonstrating the applicability of causally-guided evolution to problems in which causal knowledge is not available. Collectively, this work clearly demonstrates for the first time the promise of causally-guided evolutionary computation in a variety of forms and when applied to a range of application problems.