lexoyo / genetic-algo Goto Github PK

the current algorithm results in stupid bots

the problems are

• if the terrain is too complex the first generations never find the door and the traing fails
• there is not a fixed phenotype so I guess this is also a reason why it generates dummy bots

let's try

• a phenotype wich defines a neural network's weights (and then try to include its topology also)
• the neural network's inputs = each cell's values (empty, wall, door)
• the output = id of the command (up, down, left, right, stay)

and if it does not work, maybe try with inputs = distance of a wall up/down/left/right, angle with the door, distance of the door, previous states?

interesting links to implement a simple neural network

• a question on SO with an implementation in python in the comments (200 lines)
• a simple back propagation example