i am trying to re-create the dino game from Chrome and automatize it using neat.

The dinosaur has a fixed X position and can only jump.
The obstacles move towards the player are procedurally generated: i have a queue containing all the obstacles and when the x position of the last one has reached a certain position, a function generates another obstacle. If the first obstacle has gone too left, it gets popped, so the next one becomes the first.

The game is contained inside a Game() class. The constructor calls a reset function that creates an array for all the dinos controlled by the AI, the genomes and the networks.

def Reset(self, genomes, config):

    self.nets = []
    self.ge = []
    self.dinos = []

    for n, g in genomes:
        net = neat.nn.FeedForwardNetwork.create(g, config)
        self.nets.append(net)

        dino = Dino()
        self.dinos.append(dino)

        g.fitness = 0
        self.ge.append(g)

        self.run()

In the update loop, i iterate through each remaning dino and i add a score to them. The output of the network is processed by giving it the dino x position, the x position of the first obstacle in the Queue and a bool that stes if the dino can jump or not. I also check the collision between the dinos and the obstacles. If the dino collides, his fitness score is lowered and is deleted from the game.

def update(self):
    self.checkCollision()

    for x, dino in enumerate(self.dinos):
        self.ge[x].fitness += 0.1
        output = self.nets[x].activate((dino.x, self.obstacleQueue.queue[0].rect.x, dino.canJump))
        if output[0] > 0.5:
            if dino.canJump:
                dino.jump()

It actually (sometimes) works as it should: there are some dinos that jump trough the obstacles, but they can’t get past the second.
Then after 2 or 3 generations, they can’t get past the first obstacle.