How to fix the Markov FSM versus Neural Net ::SHE+ILA:: problem

More about How to fix the Markov FSM versus Neural Net ::SHE+ILA:: problem

See the "Markov versus Neural Net" artwork. Basically a Markov FSM had a problem with interpolating untrained prompts, whereas a Neural Net will generate a superposition, thus simulating a response which it has never experienced (creativity). We can overcome this deficiency in Markov FSMs by passing several threads through, each randomly taking different arcs at the Markov weightings table, and then forming a sum (simulating the matrix multiply feature of the Neural Net system). So each time we hit a FSM node, we take two logically independent exits, concurrently, to two parallel new states, but we limit the number of threads to two (otherwise the navigation through the tree would grow exponentially). As we approach the next pair of transitions we pick the fittest two threads out of the potential four ("Markov Darwinism"). This strategy bears a resemblance to Feynman's "sum of histories" explanation of quantum photon behaviour :)