Geoff Nitchske

An Introduction to Neuro-Evolution and its Collective Behavior Applications

3/6/2010, 12:30,
Facultad de Informática,
Aula de grados


Neuro-Evolution (NE) is a machine learning approach that combines the strengths of two biologically inspired areas of Artificial Intelligence: Artificial Neural Networks (ANNs) and Evolutionary Algorithms (EAs).  NE is the artificial evolution of ANNs using EAs, and has been shown to out-perform traditional machine learning approaches such as reinforcement learning in a disparate range of tasks that include: intelligent computer games, robot arm control, computer processor design, and rocket control. NE uses Darwinian inspired survival-of-the-fittest competition among ANNs in an artificial evolution process that leads to increasingly sophisticated solutions without the need for human design. This lecture has two main objectives. First: to introduce the design principles and computational mechanics behind various NE methods, including different approaches to encoding, recombining and adapting ANNs within the context of an evolutionary process.  Second: to discuss the application of various NE methods to solve a wide range of tasks, including complex collective behavior tasks.  Collective behavior tasks are those that require multiple agents to work cooperatively in order to solve (for example, the collective construction of a space station by multiple autonomous robots). The application of NE to solve collective behavior tasks is a relatively recent research area, and as such new NE methods for solving such tasks will be presented.

Organiza: Francisco Bellas, Departamento de Computación.