Evolving neural network. Survival agents. part II

Showcase
#1

999×526 210 KB

Regular video report of my neural network result. Added enemies and you can see, that agents starts to avoid them. On the video, agents trained about 40 min and was chosen 8 best for continue training. I was used genetic algorithm combined with neural network. Signals of agent are: 1-agent’s health, 2-closer food distance, 3-closer food angle, 4-closer enemy distance, 5-closer enemy angle; Agent can see only on his view field (gray circle).
Agent needs for increment his fitness for reproduction. It’s depends on food eaten (fitness increment) or enemy contact (fitness decrements).

P.S. It is continue of the topic “Neural Network and Qlearning. Survival agents.”
Thx.

VtVXNMB5DJE

#2

Great job:)!

The bots behave very realistic.
Did you compare your solution with other IA algorithms?

I saw at the movie that you are using four layers, input layer, two hidden and output layer.
Usually three layers are enough for multilayer perceptron neural network. Did you experiment with different number of layers?

#3

Thanks for answer :slight_smile: Yes, 4 layers just for testing, but in a future I want to allow every agent to get random layers in width and in height at beginning training and after reproduction two agents with different neural networks will create after crossing new agent with new neural network. At the moment I use just mutation of weights, but result makes me happy anyway, LOL :slight_smile:
In first version I tried to realize q-learning combined with neural network, but lack of knowledge didn’t let me do it :slight_smile:

#4

Hi

Thanks, really nice job.
Are you going to prepare a library based on your solution or just want to use it in your games?

#5

To be honest I have not thought about this :slight_smile:
Thanks again for your comment.