Transcript Lecture 12
CAP6938 Neuroevolution and Developmental Encoding Developmental Encoding 2 Dr. Kenneth Stanley October 9, 2006 Where is DE Useful? • Problems with regularities – – – – – – – – Board games Visual processing/image recognition Pictures Music Puzzles Architectures/morphologies Brains Bodies • Problems requiring high complexity – High-level cognition – Strategic thinking – Tactical thinking Where is DE not Useful? • Problems without regularity – Pole balancing • Simple high-precision domains – Very small picture reproduction • Simple control tasks – Go to the food DE Testing Paradox • DE is a mismatch for simple problems • Hard problems are too hard to just get started • Where do we begin? Progress through Benchmarks • Visualization is most revealing – Observe growth and final product Julian Miller’s French Flags http://www.elec.york.ac.uk/intsys/users/jfm7/french-flag/sld018.htm – Shapes and pictures are easy to analyze • What’s the first useful thing DE will do? Where does NEAT Fit In with DE? • If it’s a neural network, it’s obvious • If it’s just “genes,” it still works like NEAT – Need a way to get to a lot of genes – NEAT will complexify into the right number – Historical markings tell which gene is which – Speciation keeps innovation protected, incompatible individuals apart Next Class: Compositional Pattern Producing Networks (CPPNs) • Abstracting away development • A new kind of indirect encoding based on function composition • Theory and preliminary results Exploiting Regularity Without Development by Kenneth O. Stanley. To appear in: Proceedings of the AAAI Fall Symposium on Developmental Systems. Meno Park, CA: AAAI Press, 2006 (8 pages)