search

CodeReclaimers / neat-python

Python implementation of the NEAT neuroevolution algorithm
BSD 3-Clause "New" or "Revised" License
1.41k stars 491 forks source link

Neat disconnected #163

Open chill0chill opened 5 years ago

chill0chill commented 5 years ago

Hi, I am running a NEAT with more than 28 features and comparing it with a deep backprop network. I am running a regression problem.After running multiple configurations(pop size, crossover, mutation rate, connection add rate , etc). I am observing that neat is leaving out important features from the network if initial state is unconnected. If i start with fully connected network , then the accuracy is dropping a lot. Not to mention both the methods are underperforming deep fully connected network. Has anybody else faced this issue or does anybody has a thought on this problem?

mathiasose commented 5 years ago

I did some thinking about this a while back, but I didn't implement it, so take that into consideration:

It might be beneficial to create a fitness function that rewards both objective score and smaller numbers of connections and/or hidden neurons. So if two different networks perform the same at the objective task, the smaller of the two networks would get a higher fitness score than the other. This would create a selection pressure towards smaller networks, which might make the solution space exploration more efficient (assuming the task can in fact be solved by a small network).

In your case you might also want to try starting with partially (randomly) connected initial states, not just fully connected and fully disconnected.

winatawelly commented 5 years ago

Hi, I am running a NEAT with more than 28 features and comparing it with a deep backprop network. I am running a regression problem.After running multiple configurations(pop size, crossover, mutation rate, connection add rate , etc). I am observing that neat is leaving out important features from the network if initial state is unconnected. If i start with fully connected network , then the accuracy is dropping a lot. Not to mention both the methods are underperforming deep fully connected network. Has anybody else faced this issue or does anybody has a thought on this problem?

Do you ever found any solution to this problem ? i also running regression problem with 22 features and just noticed that most features doesnt even make it to the output node.

ntraft commented 1 year ago

I'm not sure what you mean by "disconnected" and "most features doesn't even make it to the output node", but it sounds like you may be impacted by the issue I filed just now: #255.