Hello, I have been working on a physics based idea, but the ultimate goal is to automate this idea with an AI solution. Basically I have a set of 10-20 input values which can change with respect to time and the computer knows nothing about, and I have a measurable success condition. i.e. A stationary creature has 40 muscles which can be contracted to produce movement in a given area of the body. The computer is able to contract each muscle by a given amount (fully relaxed 0.00 to fully contracted 1.00). The success criteria is that the creature moves forward at maximum velocity. Velocity and other such information can be fed back into the AI in real-time. Is there a particular area of AI which lends itself to this sort of trial/and error learning? There would have to be some sort of feedback in the system as the muscle contractions could not be a fixed state but would have to change over time as the centre of mass of the creature changes as it moves. There might be failure criteria such as falling over, also there might be additional criteria such as using the least amount of energy, i.e. keep muscles contracted for a little time as possible. Any help on this would be gratly appreciated, I have been looking at Neural Nets but I am not sure if they would be applicable. Thanks in advance, Mark Coleman

Take a look at Reinforcement Learning/Q Learning (see this applet for an example).

Thanks, I'll check that out.

Mark