Colloquium Details
Combinations of Evolution and Learning in Artificial Adaptive Systems
| Author: | Keith L. Downing Department of Computer and Information Sciences, The Norwegian University of Science and Technology, Trondheim, Norway |
|---|---|
| Date: | July 05, 2001 |
| Time: | 15:30 |
| Location: | 220 Deschutes |
Note: Special Time
Abstract
The crafting of everything from efficient algorithms to intelligent agents can be viewed as search in an extremely large design space. Many AI tools are useful in this endeavor, particularly those from the field of machine learning (ML). In fact, ML includes methods that go beyond learning in the standard biological sense of behavioral changes during an agent's lifetime (i.e., plasticity). To wit, evolutionary algorithms (EAs) utilize populations of agents that are subjected to the computational equivalents of "natural selection" and "breeding" to gradually improve the performance of the best and average individuals of the population over many hundreds or thousands of evolutionary generations. The last decade has seen promising results delivered by hybrid systems that combine EAs with more conventional "lifetime learning" techniques. Here, populations of agents evolve, while each agent possesses behavioral plasticity. For example, evolution may operate on a population of artificial neural network (ANN) topologies, while back-propagation or the Hebbian rule facilitates lifetime improvements of each network via weight updates. Learning has been shown to enhance evolutionary search via a biological process known as The Baldwin Effect. Alternatively, if the results of learning are encoded back into the EA agent's genome prior to mating, then evolutionary search can benefit from a form of Lamarckian evolution, though Lamarckianism has long since been disproven biologically. Either way, via Baldwinism or Lamarckianism, the combination of evolution and learning offers powerful adaptivity to artificial systems, with many hybrids outperforming their constituent evolving or learning modules. In this talk, I will a) review some basic principles and seminal research in the area of hybrid evolving and learning systems, and b) introduce my RGP system, which combines genetic programming and reinforcement learning for designing controllers for artificial-life agents.
Biography
Note: Keith is one of our first Ph.D's.