MAMEWorld Forums - EmuChat - AWESOME paper on beating Super Mario A.I. with RAM analysis

Take it for what it is, not for what it could be. The idea is novel, yes, but fascinating none the less.

I agree he's a douche...

> >
> http://www.geek.com/games/computer-program-learns-to-play-classic-nes-games-1552024/
>
> It's not as cool as it sounds. To boil it down to the most basic features, the author
> has created a program that:
>
> a) Evaluates the entire 2048 byte chunk of memory of an NES system without any actual
> understanding of what any of it means, only that some values can be seen to go "up",
> and "up is good".
>
> b) Runs through an emulator, trying random inputs and finding the sequence of inputs
> that increases the memory locations for which it was already determined that "up is
> good" the most
>
> c) Uses some modicom of smart heuristics to reduce the search space (instead of
> trying all 256 possible input combinations every frame, have a better algorithm for
> only evaluating those inputs that are most likely to be good)
>
> d) Uses some other mechanisms that I didn't read all of the details about to try to
> find workarounds to the inevitable shortcomings of this simplistic approach
>
> e) After chewing for a long time comes up with an input sequence that was able to
> raise the important memory location values the most
>
> It's not a sophisticated A.I. by any stretch of the imagination. It's a simple search
> function with some tricks to try to make it more effective, combined with a novel
> (although fairly simplistic) mechanism for evaluating the search function (i.e.
> trying to find values in memory that can consistently be made to increase).
>
> It takes an hour of CPU computation just to output the inputs for 16 seconds of
> gameplay. It cannot play the game real time. Furthermore, the program doesn't
> actually "exploit bugs in the game", in the sense of knowing about them and
> anticipating and using them. If a random input sequence at a specific moment of the
> game happened to result in Mario surviving due to a game bug, then so be it. The A.I.
> will not use that information to help it plan a future move.
>
> Additionally, the author somewhat self-aggrandizes the program with the long "white
> paper" title and his video. And the white paper itself is more like an informal email
> describing what he did and then put into the shape of a white paper, than an actual
> white paper meant to be taken seriously. I've never seen a white paper use the term
> "horse shit" before (nor the dozens of other informalities it also uses), for
> example.