'Neural Network Trainer: Fitness & Crossover'

01:32 Sep 11, 2022
'Just another small update to this project -- Implemented a first pass of the crossover (breeding, mixng/matching genes & mutations) and fitness (handles how a score is determined for each game) panels.  In this video I just show how setting the fitness component values can impact the resulting training behavior of the agents.  The Dot Product score is simply a measure of how closely an agent\'s trajectory matches the direction of the target.   At first, it is trained to maximize this fitness measure (move towards target location), but when \"Bigger is Better\" is unchecked, that means a lower score for this fitness component is considered good, which means the agents will be trying to go in the opposite direction of the target.  Finally I switch it back and show how the agents re-train to going Towards the target.' 
See also:

comments