I am trying to develop neuro controllers by using genetic algorithm to train a neural network of fixed topology. An issue I am facing is objective function for motion of robot from a source to goal . Parameters affecting the function are angle towards goal at any given location , distance from goal at the end of T steps (T is the number of steps for which each member of population runs) and finally obstacle avoidance . The issue I am facing is the weights with which all these parameters (which are basically penalties) must be combined as in my experiments I found each parameter does not affect the cost in same manner so if the function I take is

f(x)= a*P1 + b*P2 + c*P3

where P1 P2 P3 are penalties due to parameters under consideration how I should formulate or if necessary learn a b and c.