AI Master's thesis

[KHRZ]

Although I won't be writing it untill next fall, I have to register a contract of my Master's thesis and supervisor in December, meaning I pretty much have to come up with something ASAP, so I can pick the relevant supervisor, and not have the risk of being passed up by that supervisor if he is too popular.

I have some interest in trying to find mathematical proofs, I have this snobby math student friend that don't think AI can do any intelligent "analysis", and challenged me on the Pythagorean Theorem. So I showed him (on paper) a system that would quickly find a popular geometrical proof of the Pythagoren Theorem. (Less than 100 tries of finding a geometric configuration from which the proof is obtainable, if it could rate the relevance of where to look first well. The whole idea is that it can look at very, very many candidates, like a real person, but still judge the best place to look.

But actually making the system is a massive job, it would contain both semantic knowledge networks, heuristic searching in first order logic with some extensive math vocabulary, translation of this first order logic to aquire math expressions to look for the proof, and have Wolfram-alpha level math processing to solve the proof equations. So this is a massive system, probably way over my head. I would love to see it actually work and find some proofs in this way (I know there are already plenty of provers that finds more general proofs and whatnot though. But they often use exhaustive proofs on abstract theorems etc, so IMO, this is more of the "analysis" way of finding proofs).

I am still considering if I should present this idea to a professor. Any thougts?

I've looked at some of the earlier Master thesises, and it's common to just write some AI that plays some game, which may be what I end up doing in the end. Now I was thinking, is there anywhere in Dolphin it would be possible to insert some AI? (As in not just looking at the screen of the game, maybe some game have relevant datas in memory for the AI to use that makes it easier to write the AI.) If anyone have looked at specific games, or know in general if this sounds feasible, please tell.

[delroth]

https://www.cs.cmu.edu/~tom7/mario/mario.pdf is probably relevant to your interests.

[KHRZ]

Ah yes I heard about that AI a while back. Now it does use "hundreds of (NES) CPU hours", I assume for one playing session of a game, and the same approach in Dolphin would be a bit slow. What could possibly work and run in real time, is to try to build predictions of actions effect on the game world, rather than looking ahead in time at the actual consequences.


What I was thinking more of though is either write an AI for playing a specific game, by detecting various variables in that game, and if there is good access to data like world geometry, character locations, etc. But it might make more sense to do as a game hack than being related to Dolphin for some games (like Super Smash Brothers, I think AI tweaking have been done already in project M).

One typical task is to build a case-based reasoner for a game, and try to train it to play well, the whole working-on-the-memory-with-no-knowledge idea is interesting, but I don't think the time has come yet for it to work out very well (in the future, maybe an extremely good AI will read the memory and understand the game, like a person understands from looking at the screen)