I made that project from scratch in javascript which combine genetic algorithm and neuronal network. Creatures have to eat food to survive, they start without any knowledge and thanks mutation they will start to learn and evolve by themselves.
Genetic algorithm is here to replace back propagation in neuronal network. We let our creatures correct the network themselves.
I was facing a problem with cross over, in fact, in some videos on the internet which try to do the same things, they don't really use a cross over with their neuronal network because in fact, two solutions can be so different and melt them together won't help to make progress.
I think that's why my version 0.03 and below won't progress easily. Because it's random and we loose the goal of the genetic algorithm. After some research, I found NEAT, which seems to resolve my problem.
Here is a link about NEAT : http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf
When I started that project I was a bit lazy about sensors. First of all I was given the relative coordinate of the closer food item and his own rotation as input for my neuronal network. In theory that's enough information for my creature and he should "eat" food. But that didn't work like that. Indeed, I waited more than 1000 generations and nothing happened. For me, it is because rotation is an input and an output which make my network more complex because creatures had to combine input together to obtain a good result.
That's why I decided to change input by putting sensors for my creatures. Now, each creature has 5 sensors which can detect food if there isn't food, sensor return 1 else they return a number between 0 and 1 according to the food distance. You can find find the sensor code in sensor.js.
Here is a schema before and after the changement :
Genetic art : https://github.com/dylandoamaral/genetic-art
- Best code organisation
- Add a test zone for best creature and some chart
- Comment my code
- Import with javascript ES6