The gBrain Project

[feha]

Will it save "sessions" in a txt file? That would allow you to develope the ai without having to worry about everything being gone by a simple click of a button (or mingecombat).

[Lyinginbedmon]

Storing runtimes would be a simple matter of storing the gene strings in a txt file, yeah. The central core could easily be coded to save the latest strings every few minutes.

Which still means having both sides output their modified gene strings (Which also theoretically means we can network a bunch of gBrains together, which is kinda awesome) so that the central core can save them and simultaneously update the strings it's outputting for them to work from. Alternatively we could just have each side save it's genes but I think the benefits from the above outweights the code save of that.

[Lyinginbedmon]

One of the larger considerations for the gBrain on the Intuition side is how to tell it to attach a certain number or proportion of it's neurons to a given action and to which sensory inputs those actions are associated. This could be coded into the genes but it's a little tricky to consider.

I think it'll best as a set number of neurons, then a percentage of those neurons assigned to a given action. Then, we can use a sensory hub to collect the inputs and sort them into the relevant arrays for the neurons to work through.
Round the neuron total for each action down to the nearest even number, then halve the number of neurons at each level to eventually work down to a single neuron ("After all this thinking, do we or do we not do this?" basically).
The Neur array is temporary, so we can easily remove the entries as the loop runs through it to ensure a given neuron is not attached to two different actions, preventing neurons from bleeding through into other thought processes.

(Side note: This means each thought process must have a number of neurons equivalent to 2^N, where N is the rounded down log2(X) and X is the initial number of neurons, and log2(X) is also the number of iterations of thought)

[feha]

My idea was not save every x minut (though might be usefull against crashes), but actually on a chat cmd and last().

How about giving its location in a big gvar (or another way that makes it easy to get pos of all "brains")? Then you feed it stuff like traces (pathfinding), target and whatnot. But for target, you can just make a check in the global and find the closest "brain". Aka a way to run a findByClass("brain") except there is not really anything of that class.

[Riddler]

Not to be a killjoy, but ither way your looking at a considerable ammount of training time no matter what you do. Quite possably more than would be practical to give.
Having gotten that out of the way. I hope this works, definatly give it a go

One problem I can see with this is the problem of training, from what I know (which is by no means 100% accurate) to train nodes you give them a value and then compare what is known to be the real awnser with the awnser given by the node then ajust acordingly. This is fine for math problems and some other things however for controlling something, not so much. The goal for example, could be to lift its arm. The problem being the AI wont just know it needs to find a way to lift its arm it needs programed to know it should lift its arm. Also it will need to know how to tell if its lifted its arm in order to know it has acomplished its goal, so it needs to be programed to know set X to Y(node output) until Z is acomplished. Ither way hours of codeing will need to be spent ither making an expression to controll the drone or telling the drone the what it needs to do and how to know if its done it and then training it to do it.

While it may be possable to programe it to learn new objectives and how to know if its acomplished them, thats a huge ammount extra training time.

[Schilcote]

I think that once you get the robot familiar with what it's supposed to do, you can usually leave it mostly alone and let it simply train itself (that's the main point of neural networks and learning machines in general, isn't it? They can learn from experience, effectively training themselves...).

[Lyinginbedmon]

That's why the intuition side has basic definitions, but the pattern side undertakes the chosen action. Intuition just has the neurons that gradually respond to the stimuli better, Pattern discerns what an action actually consists of. It's the "fine tuning" part of the brain where the Intuition is just that, instinct and the basics necessary for function.

[SystemsLock]

Glad to see your using my neural network idea. It sounds like we really share a passion.

Still, I doubt this is possible in a practical sense. You've simply got too many variables to account for in a drone. Neural networks work better in theoretical environments with more constant and direct cause/effect scenarios. Your going to have a hard time getting enough training data and training the network in any reasonable amount of time.

[steak knife]

This project definitely intrigues me. I wonder what sorts of mental modules you will have to use in order for it to have some sort of instinct. In other words, what will be your starting point? won't the brain require some sort of place to begin with respect to it's senses and motor functions?

[SystemsLock]

It really depend on how he wants to do it. Traditionally speaking no, neural networks do not have any built in functionality. But if he opts to use them as more of a decision aid then an artificial brain, he may have to have many processes automated prior.