omreko at October 1st, 2005 04:46 — #1
my qustion about using genetic algorithm for already trained ANN
i have developed ANN for actual system, i have the input patterns and Target patterns, if i need one target i forward the ANN with iterative inputs and check the ANN output for the required target. i need to implement genetic algorithm to select the proper input combination to achieve the required target quickly without too much iterations, can Genetic algorithm do that?
geon at October 1st, 2005 06:44 — #2
Probably not without "too much iterations". It takes a lot of generations and a pretty large population to achieve anything usable.
alex at October 1st, 2005 07:24 — #3
How many is "many iterations" in your case? What kind of population size, etc?
If you have specific input and output patterns for all cases available you might want to try using backpropagation. Training the network genetically gives good results if you don't have specific input and output patterns but if you actually simulate the agent and its surroundings using the network and afterwards assess the quality of the agent within the given settings using rather abstract assessment.
So if you need an agent that behaves well in a certain simulateable environment then genetic training might be the way to go. If you want specific "answer" signal patterns given a specific input pattern (for example training the network to do a xor) then other methods might be better suited.
nomadrock at October 1st, 2005 10:45 — #4
The easiest way is to simply analyze the network. What inputs from the last hidden layer will produce the desired output? Then what inputs to the next layer will produce the values in that layer. You keep doing this until you hit the input layer. Then find any solution to your system of equations and you have it solved Note for the general equation there are likely very many solutions, but you have to restrict it to only allow values between 0..1