CC AI

[tensorpudding]

Oh, you're talking about game monster AI. I thought you wanted to create an AI which could be trained to solve CC puzzles in reasonable time. That would probably be rather hard.

[Lessinath]

Either one - or both!

[lotu]

Oh, you're talking about game monster AI. I thought you wanted to create an AI which could be trained to solve CC puzzles in reasonable time. That would probably be rather hard.

We got off topic, I am taking about an AI to control Chip and solve puzzles as a player would. You are correct that it is a rather hard problem. We are going to try and put the same restrictions on the AI that a human would have, (i.e. only see a 9x9 window), as this is less about making a solver and more about learning about AI.

[geodave]

Will the AI know things like "Oh, this is a Tyler level, so I should look for misleading item-swapping and towers of blocks on ffs?". Or "pieguy made this level, so I should just give up and skip it?"

[Hornlitz]

LOL @ Geodave

also, that does bring up a point... Insane levels. Would it say "Oh, this level is impossible." even though it may be?

[Lessinath]

A really well programmed AI would try random things that should do nothing but won't kill it if it can't figure out what to do.

[tensorpudding]

Without knowing a ton about AI, having just skimmed some books on machine learning, I would say that those strategies that Dave brought up would be really hard to teach an AI. How do you describe a puzzle that looks like a Tyler puzzle in a way that a machine can reliably determine the difference between it and some other puzzle that isn't a Tyler-style puzzle, but uses similar elements? The same for determining solvability...I doubt that is actually possible without brute-force search.

[Hornlitz]

Without knowing a ton about AI, having just skimmed some books on machine learning, I would say that those strategies that Dave brought up would be really hard to teach an AI. How do you describe a puzzle that looks like a Tyler puzzle in a way that a machine can reliably determine the difference between it and some other puzzle that isn't a Tyler-style puzzle, but uses similar elements? The same for determining solvability...I doubt that is actually possible without brute-force search.

You could have it just ask if, say, it's a Tyler level or if it's and Insane Level...

[lotu]

Tensorpudding is right, the AI isn't going to be able to understand the concept of the author of this level, tends to this so I should look for this type of solution. It doesn't even understand the concept that levels have authors. Though in principle you might be able to you neural networks bias how you explore the level, and then train that off existing levels. But I doubt it would work in practice. I suspect if you would have to be able to carry on a deep meaningful conversation about politics before you'd see an AI being able to make that kind of deductions.

However, I have progress you can see the the AI running on this level here. As you can see the AI doesn't get the know where the walls are when it starts.

http://www.youtube.c...bed/_yc0EpqKdi8

[Lessinath]

That's a great start. Can it handle keys and locks yet?