Artificial evolution addresses difficult, open-ended problems
We focus on two such problems:
- The problem of developing strategic behaviors in vision-based agents for virtual world tasks;
- The problem of generalized control strategies for the AOTF "smart filter" camera
The performance of artificial evolution hinges on the choice of the representation of points in the target search space (i.e., candidate solutions) and the search operators which act upon them. The traditional approach to this issue is empirical design and testing of representations and sets of operators for use in particular domains. Because this design process may need to be restarted from scratch when arbitrary aspects the target domain change, the alternative that we favor is to simply permit aspects of the representation and search operators to themselves evolve (an approach known as "self-adaptation" in the evolutionary computation literature), as is implicit in natural evolution.
Drawing upon research into self-adapting mutation rates, past work has included analysis of how evolution acts upon search-level traits (i.e., traits which affect the representation and/or operators) in bitstrings, and how information can be extracted off-line from traces of artificial evolution runs (evolving weight-vectors for feed-forward neural networks for face recognition). in order to enhance further search in the domain. Current research focuses on how to adapt to a search space -- currently, the the weight space of fixed-architecture recurrent neural networks controlling visual agents in a virtual world -- during evolution so as to promote improved evolution in this space in the future.