Druchii.net

Given enough time to learn, the program might try to take over the world some day, but that's a risk I'm willing to take.

Looks interesting...Just don't let it on the internet! It will teach it bad thing!

Saintofm wrote:Looks interesting...Just don't let it on the internet! It will teach it bad thing!

Didn't the last AI test on the net become racist and genocidal within 24hrs?

http://www.bbc.com/news/technology-35890188

Amboadine wrote:

Saintofm wrote:Looks interesting...Just don't let it on the internet! It will teach it bad thing!

Didn't the last AI test on the net become racist and genocidal within 24hrs?

http://www.bbc.com/news/technology-35890188

That's the one. Heaven help us if the next one gets on an MMO. That will truly start the robot uprising.

As for the OP: I think its interesting what they are doing and hope it does some good but some of the algorithms kinda went over my flu addled mind at the moment.

Amboadine wrote:

Saintofm wrote:Looks interesting...Just don't let it on the internet! It will teach it bad thing!

Didn't the last AI test on the net become racist and genocidal within 24hrs?
http://www.bbc.com/news/technology-35890188

This AI program will be released on the internet, but it won't have Twitter. So we're safe...
... I think. Anyhow, let's not squabble over potential risks and look at the possibilities!

Oh absolutely. Possibilities are extremely exciting.

I think that, once you have it trying out pile-in moves for multiple models, it's going to take many (many, many) generations to come up with even one legal solution (i.e. each model ends up closer to their nearest enemy and within 1" of a friend) unless you constrain it from the start. To put it another way, the AI should probably only concern itself with choosing between legal pile-in moves, not just moving every model randomly. Otherwise, you're going to waste a lot of processing time.

That's an interesting point. I had the same concern in mind, that the eventual pile in strategy should be legal. I think we have to allow them, at least as intermediate results.

Perhaps the solution to a better pile in strategy may need a side step right outside the box. One model could move too far from the group but in doing so, could reveal a new path and strategy al together. It would be sad to lose this before it can grow into a solution.
Preventing illegal moves would be difficult as well, demanding complex analytics to calculate the "remaining moves" for the final models, which are heavily bound by the moves of the first models.
If we'd adopt a try first and fix logic, then the process to generate a new pile move could take a while as it may have to correct itself multiple times. This is kind of an evolutionary algorithm in itself. So why not use the general program for it?

My guess is that the "natural" solution would be to penalize illegal moves during the fitness test. This will make the algorithm favour legal solutions, but still think outside the box when the legal solutions presented aren't very brilliant.

Come to think of it, this would make a great case study. I'll make this a parameter so we can arbitrarily define how heavily an illegal move is penalized. That way, we can see how big the impact is on the creativity of the A.I..