schultz@halley.est.3m.com (John C. Schultz) (03/14/91)
I am trying to use a C code representation of a back prop neural network as the evaluation function for a genetic algorithm optimization routine based on the GENESIS 1.2ucsd code. While the concept works quite well (after a little hacking at the interface), some optimized solution are values less than 0.0000. Physically the output variable being optimized is a thickness and thickness' less than 0.00 are meaningless, at least to me. I doubt that I can limit the neural network values (I assume that the GA has found some strange, untrained portion of the NN to exercise?) However, can I limit the output values of the GA to be positive? I would be quite happy to get a set of test points which results in a ouput variable near 0.00000 (but positive). Thanks for any suggestions. -- John C. Schultz EMAIL: schultz@halley.serc.3m.com 3M Company, Building 518-01-1 WRK: +1 (612) 733-4047 1865 Woodlane Drive, Dock 4, Woodbury, MN 55125 How to include the taste of Glendronach in a multi-media system?
esrmm@warwick.ac.uk (Denis Anthony) (03/15/91)
In article <SCHULTZ.91Mar13223859@halley.est.3m.com> schultz@halley.est.3m.com (John C. Schultz) writes: >I am trying to use a C code representation of a back prop neural >network as the evaluation function for a genetic algorithm >optimization routine based on the GENESIS 1.2ucsd code. > >While the concept works quite well (after a little hacking at the >interface), some optimized solution are values less than 0.0000. >Physically the output variable being optimized is a thickness and >thickness' less than 0.00 are meaningless, at least to me. > >I doubt that I can limit the neural network values (I assume that the >GA has found some strange, untrained portion of the NN to exercise?) >However, can I limit the output values of the GA to be positive? > >I would be quite happy to get a set of test points which results in a >ouput variable near 0.00000 (but positive). > Are you using total output error (or similar) as the cost function for the GA ? You could adapt the returend error by (say) giving a high cost when the result is meaningless (output < 0).