训练神经网络用遗传算法玩飞鸟 - 为什么不能学习？

Question

I have been learning about neural networks and genetic algorithms, and to test my learning, have tried to make an AI that learns to play flappy bird:

I have left it running for at least 10 hours (overnight and longer), but the fittest member still fails to show any significant advancements in intelligence from when I began the simulation apart from avoiding the floor and ceilings. The inputs are the rays (as you can see above) that act as sight lines, and the network is fed in their lengths, and the birds vertical velocity. It seems that the best bird is essentially ignoring all the sight lines except the horizontal one, and when it is very short, it is jumping. The output is a number between 0 and 1, if the output is larger than 0.5, then the bird jumps. There are 4 hidden layers, with 15 neurons each, with the input layer feeding forward to the first hidden layers, then the 1st hidden layer feeding forward to the 2nd one ... and the final hidden layer feeding forward to the output, the dna of a bird is an array of real numbers representing the weights of the neural networks, I have made another project using the same style of neural network, and genetic algorithm, in which ants had to travel to food, and it worked perfectly.

Here is the code: https://github.com/Karan0110/flappy-bird-ai

Please say in the comments if you need any additional information

Please can you say whether my method is flawed or not, as I am almost certain the code works correctly (I got from the previous working project).

Answer 1

I like your idea, but I suggest you change some things.

• Don't use a network with a fixed structure. Look up Neural evolution of autgmenting topologies and rather implement it yourself, or use a library like neataptic.

  • I don't believe your network needs that many inputs. I believe 3-5 sensors (20-50° gaps) would be enough, since many of the input values seem to be very similar.

If you are not sure why exactly your project is not working try this:

• Try view an image of your current best network. If the network doesn't take important sensors (like the velocity) into account, you'll see it instantly.

• Make sure all of your sensors are working fine (looks fine in the image above) and be sure to nkrmalize the values in a meaningful way.

• Check if the maximum & average score increases over time. If it doesn't your GA isn't working properly or your networ receives inputs that are not good enough to solve the problem.

One trick that helped me out a lot, is to keep the elite of the GA in a seperate array. Only replace elite networks if some other network has performed better than the elite. Keep the elite trough all the generations, so once your algorithm finds an extraordinarily good solution, it won't be lost in any future generation if nothing else performs better.