Could we handle an AI?

[Nikita]

Apart from the fact that it would probably take more combined E2 ops and total entities than a beefy gaming pc can support, when each neuron or a group of neurons is presented by a chip, you could make an automatic wirer that links and unlinks neurons to conduct learning.

[Lyinginbedmon]

Well if we get genetic algorithms into the mix, then each neuron is the same repeating system modified by values derived from the gene string. In which case, one E2 chip could run through a whole bunch of neurons by looping the system through the string.

The trick then is to have the chip run through states. The initial state is the inputs available run into the first hidden layer of neurons, the second state is then the next layer of neurons running through the reactions of the first (Stored in an array most likely)

Think it's probably about time I made a thread on gBrain, to give us something to consider on a realistic level...

[Neo Kabuto]

Well, I got my little experiment working. ^_^

Since it's such a simple goal (a NOT gate, which doesn't give me much feedback on how close I am to the correct weights), I couldn't use a genetic algorithm, and instead used a quasi-random search. The bottom part is just the single neuron required for the gate and the input button, and the top part is the "Learning Module". It's not exactly finished or optimised, but it does figure out that the solution requires a negative weight pretty quickly.

[Pangamini]

Why interval? Call the chip by inputs

[Neo Kabuto]

Why interval? Call the chip by inputs

I think my reasoning behind that was that it only would activate then when something changed, but a logic gate only has two possible outputs. Interval would let me keep it changing without the gate's output or input changing. This is too simple a problem for genetic algorithms, like I said before, so this just keeps plugging weights in until it works.

[Pangamini]

Yes but you might make one clock input for neurons so they are time synchronized

[Nikita]

I think runOnTick(1) is a more natural way to make sure E2s execute at the same time, based on the fact that normal wire can only be updated each tick.

[Neo Kabuto]

Yes but you might make one clock input for neurons so they are time synchronized

I think runOnTick(1) is a more natural way to make sure E2s execute at the same time, based on the fact that normal wire can only be updated each tick.

Those are both good ideas. I need to remember them for next time.

[Lyinginbedmon]

Posted a thread on the gBrain project, think it should be a good practical exercise for us outside all this theory.

[Nikita]

Random rambling #53: Let's say I wanted to make an AI learn by silently listening to what people say. Does anyone have any idea about a good volume of text that I could feed to it? Something that you can actually learn from, i.e. less dragons/spacecraft descriptions and more things like "tree was struct by lightning and fell down".