Evolving Finite State Behavior using Marker-Based Genetic Encoding of Neural Networks (1991)
The ability of natural organisms to process sensory data arose in response to evolutionary pressures. Simply put, those organisms having a better knowledge of their environment can react to external changes more rapidly and with greater precision, thereby gaining an advantage over less capable organisms. This project attempts to apply suitable evolutionary pressures to effect the development of sensory processing capabilities in artificial organisms. In pursuit of this goal, key biological concepts are used to guide the simulation. A genetic algorithm is used as the mechanism for development, neural networks are used to implement processing function in the artificial organisms, and sensory input is provided in a raw, unprocessed form. Our experiments show that by offering a survivability advantage to those organisms which are best able to use their sensory input, neural networks quickly evolve with the capacity to process sensory (visual) input effectively. The development of the complex neural network structures needed for this task is facilitated by a new gene-to-neural-network mapping which allows networks of arbitrary complexity to evolve.
View:
PDF
Citation:
Technical Report HR-91-01, Department of Computer Science, The University of Texas at Austin.
Bibtex:

Brad Fullmer Undergraduate Alumni