MLP Neural Network extension

I'd find it very helpful to see some more intricate examples using this object. I've searched and looked all over but I can't find any... (that is, aside from the very basic example that comes with the plugin)

(sorry for the bump)

This extension looks like it could be really handy. It's too bad that it seems the developer isn't around to make it anymore.

I just LOVE neural networks!

I'm considering doing some serious testing with it. Have to think of a game that is simple enough (in terms of states that the inputs would need to capture, and of actions that a player would be able to execute), and also something that happens in steps within a logical data structure (for example, it happens in turns instead of realtime). This way we would be able to train the NN with a batch, using a fast loop to instantly teach by alternating between hundreds of varied cases instead of training 100 times with the same data (which can lead to pseudo-expert behavior that only for the last presented case, making the NN 'unlearn' what it already knew).

Damn, I'm almost out to start this right freaking now! I just hope it's not too late for this extension to shine!

Actually, it seems that this extension has a final version released in another thread (where you can download said latest version), so I don't think it's too late.

By the way, NecroLuigi, do you think you could make an example for the Neural Network object? I would really appreciate it if you could.

Quote from happygreenfrog

By the way, NecroLuigi, do you think you could make an example for the Neural Network object? I would really appreciate it if you could.

I will surely do that as soon as I completely understand what is going on!

Basically, the first thing is identifying a task that wouldn't be easily done through traditional algorithms, so that the usefulness of neural nets in comparison to any other approach is made evident. I think I'm going with recognizing handwritten digits! Next step is to identify what are the relevant inputs and how to standardize them (information that needs to be gathered and treated to be fed into the network), then putting together a bunch of training cases (a set of values to the inputs tied to a specific, correct answer), and then training the network - that is, exposing it to these training cases many times and backpropagating the weight adjustments, something the extension is supposed to do in its training function). Finally, setting up an environment for the user to 'play' with the neural net - probably a surface where he can draw a digit and see if the neural net can correctly guess what digit that is.

I'm not sure how difficult that would be cause I'm not used to the extension and don't know how some of its internal details were set (ex. are the neurons using a sigmoid function to standardize its values?). I'll mess around with it a little bit and see what I can do!