REU Site on Integrated Machine Learning Systems:
Evolutionary Computation


Dean F. Hougen


  • Population size and diversity in genetic algorithms (Hougen). While evolutionary computation methods are capable of providing outstanding results on real-world problems (as illustrated by the Human-Competitive Results awards of the Genetic and Evolutionary Computational Conference), there is still much research to be done before these methods can be used in integrated applications. Evolutionary computation methods typically require users to set a large number of parameters and the behavior of the system may still be influenced by other factors (for example, resilience, Thomason and Soule, 2006). One factor that has often been credited with differences in system performance is diversity of the initial population. Recent research suggests that population size is a much more important factor than initial population diversity (Diaz-Gomez, forthcoming) but the full answer is not yet known. We investigate the effects of initial population diversity and diversity retention in a variety of real-world applications.

  • Maximizing Fitness while Minimizing Trial Time. Evolutionary computation methods such as genetic algorithms and genetic programming proceed by alternating between generating candidate solutions and evaluating their fitness. When using these methods for learning control of physical systems, evaluating fitness often consists of running the system through tests of its abilities, known as trials. The accepted wisdom seems to be that one should run several trials per candidate solution (Moriarty, Schultz & Grefenstette, 1999). Unfortunately, this is an expensive undertaking---whether learning using actual hardware or high-fidelity simulations, trial time will almost surely be the limiting factor in learning. Fortunately, we have found reason to question this accepted wisdom. In a surprising result we have found that, for one benchmark problem at least, minimizing the number of trials per fitness evaluation resulted in faster learning (that is, a faster increase in fitness per trial run) (Hougen, Carmer & Woehrer, 2003). This project explores the relationship between trials per evaluation and fitness increase over a range of additional benchmark problems and testing conditions.

  • Memetic Learning in Real Robots. Memetic learning is a novel machine-learning method that combines components from group-level learning methods (including various forms of evolutionary computation) and individual-level learning methods (such as reinforcement learning). Memetic learning is based on the way humans learn through imitation of successful behavior they observe. Imitation should allow systems to learn more quickly than they would if they were restricted to learning based solely on their own experiences. This is because they have the experiences of others to guide their own learning. This is particularly important in learning from data in real physical systems, such as robots, because experiences are so expensive to collect in these systems. In simulations of control problems memetic learning has proven quite effective at learning with fewer experiences than similar evolutionary computation methods (Hougen, Carmer & Woehrer, 2003; Eskridge & Hougen, 2003ab).

REU Home
Back to Projects

fagg [[at]]

Last modified: Wed Feb 20 00:03:48 2008