this result leads me to ask,

*what resolution of weights is sufficient?*having such a large degree of resolution in the weighting creates a huge solution space. to see this, imagine the weights of this network are limited to take on only 10 different values and there are only 2 weights in the nn. the solution space is the number of possible configurations. there 10 possibilites for the first weight and 10 possibilities for the second weight. thus there are 10*10 or 100 possible solutions. the aispy nn has 51 weights, so constraining the weighting to 10 possibilities would have 10^51 solutions, a big number but not nearly as big as the solution space now stands. the floating point number scheme in java is represented by 64 bits. this gives 16 places past the decimal. the genetic algorithm favors searching weights in the low single digits, between -3 and +3, say. so as things stand that solution space is verry big ~ 6*10^16^51. this is a stupendously large number 6*10^816 or many hundreds of orders of magnitude greater than a googol (10^100.) on one hand, it is testament to the power of the evolutionary process that it can find any profitable solution in such a huge space. on the other hand, this big a solution space is overkill and i think is a situation whereby more might be done with less. i think i will have to revisit this problem with a reduced solution space.