News Ultimarc Emu Loader Programming MAME Artwork progetto-SNAPS MAMEUI HBMAME EmuChat Hardware The Loony Bin MAME Artwork X

Entrance | Main Index | Search | New user | Login | Who's Online | FAQ | Rules

MAMEWorld >> EmuChat Index   Flat

Bryan Ischo MAME Fan Reged: 03/28/10 Posts: 358 Send PM Re: Okay Debbi Downer... 04/23/13 08:36 AM Reply > I think this misses the point a little. It's not about writing an AI to play a game, > clearly there are much more effective ways to do that. This paper is about writing a > program that can *learn* to play a game, which is a rather different and much more > difficult task. Except the program doesn't "learn" anything. It just tries different random inputs every frame over and over again and then keeps the ones that gave the best results. It has some heuristics for trying to guess the best next input and also for backtracking to a previous state and starting over again when it thinks it can do better. That's basically it. In the end you get a canned set of inputs that it determed were the best. If you feed it a different starting point then it will do the whole thing over again without having "learned" anything about what the best choices to make are. And it takes an hour to generate 16 seconds of best-case random-button-mashing. The author even plays back a segment of one run and talks about the actions of Mario like the game "meant" to do what it's doing. Mario is running right and tosses some fireballs just at the right time to hit some Goombas that haven't even appeared on screen yet (when they do appear, the fireballs are there just in time to hit them). The game didn't "plan" this move; it just so happened that when trying one out of the thousands of input combinations around that point in the game one of them happened to be sending a fireball at the exact moment that was going to end up corresponding to a hit on a Goomba. The playfun algorithm then detected that if it saved that particular input sequence, it resulted in a higher score. So it was kept. It's kind of just randomly mashing buttons over and over again and then keeping the "luckiest" sequence. Not that the result isn't cool - if you can wait an hour to enjoy 16 seconds of really lucky button mashing then it's fun to watch. But the thing is in no way "learning" or trying to model any kind of decision process. It's just searching through button mash sequences and finding the best one.

Entire thread Subject Posted by Posted on AWESOME paper on beating Super Mario A.I. with RAM analysis italie 04/15/13 02:07 AM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis Bryan Ischo  04/22/13 09:53 PM Okay Debbi Downer... italie  04/23/13 02:11 AM Re: Okay Debbi Downer... Bryan Ischo  04/23/13 02:27 AM Re: Okay Debbi Downer... DMala  04/23/13 07:27 AM Re: Okay Debbi Downer... Bryan Ischo  04/23/13 08:36 AM Re: Okay Debbi Downer... Anonymous  04/25/13 04:53 PM Re: Okay Debbi Downer... R. Belmont  04/23/13 08:05 PM Re: Okay Debbi Downer... RetroRepair  04/25/13 03:06 AM Re: Okay Debbi Downer... DMala  04/25/13 06:26 AM Re: Okay Debbi Downer... Tomu Breidah  04/23/13 05:00 AM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis DMala  04/15/13 05:10 AM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis R. Belmont  04/15/13 06:14 PM Neat try, but not really comparable. Firehawke  04/16/13 04:40 AM Re: Neat try, but not really comparable. Anonymous  04/22/13 12:01 PM Re: Neat try, but not really comparable. GatKong  04/23/13 08:38 PM Re: Neat try, but not really comparable. Bryan Ischo  04/24/13 09:50 PM Re: Neat try, but not really comparable. Anonymous  04/30/13 07:45 PM Re: Neat try, but not really comparable. Olivier Galibert  04/30/13 03:23 PM Re: Neat try, but not really comparable. Anonymous  04/28/13 12:34 PM Re: Neat try, but not really comparable. R. Belmont  04/25/13 04:52 PM Re: Neat try, but not really comparable. Vas Crabb  04/26/13 03:31 AM Re: Neat try, but not really comparable. Anonymous  04/28/13 12:37 PM Re: Neat try, but not really comparable. italie  04/28/13 03:55 PM Re: Neat try, but not really comparable. Anonymous  04/29/13 04:12 PM Re: Neat try, but not really comparable. R. Belmont  04/29/13 05:07 PM Re: Neat try, but not really comparable. Anonymous  04/29/13 06:23 PM Re: Neat try, but not really comparable. Vas Crabb  04/29/13 07:26 PM Re: Neat try, but not really comparable. Anonymous  04/30/13 03:30 PM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis Stiletto  04/15/13 08:25 PM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis Anonymous  04/15/13 11:16 PM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis Stiletto  04/15/13 11:27 PM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis R. Belmont  04/16/13 05:28 PM Re: AWESOME paper on beating Super Mario A.I. with RAM analysis Tomu Breidah  04/15/13 04:26 AM

Extra information Permissions Moderator:  Robbbert, Tafoid  0 registered and 37 anonymous users are browsing this forum. You cannot start new topics You cannot reply to topics HTML is enabled UBBCode is enabled Thread views: 2646

Browser Test | Delete Cookies atom-guru atom-guru atom-guru atom-guru UBB.threads™ 6.3.2