Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Spehss _]

I am watching an AI learn to play Mario through trial and error and I find myself getting more emotionally invested than when I watched any let's player ever. When it jumps on a koopa and kicks the shell at another koopa, it feels great. When it messes up and runs past 3 mushrooms before falling in a pit, I want to yell at it to go back and try again. Like watching a horror move protagonist go back to look for the cat and you know they're going to die and you want them to stop being stupid and go back.

[Sergarr]

I am watching an AI learn to play Mario through trial and error and I find myself getting more emotionally invested than when I watched any let's player ever. When it jumps on a koopa and kicks the shell at another koopa, it feels great. When it messes up and runs past 3 mushrooms before falling in a pit, I want to yell at it to go back and try again. Like watching a horror move protagonist go back to look for the cat and you know they're going to die and you want them to stop being stupid and go back.

I wonder what sensory input that AI is getting, because it appears to be not enough to prevent Mario from repeatedly running into Koopas.

[Spehss _]

I wonder what sensory input that AI is getting, because it appears to be not enough to prevent Mario from repeatedly running into Koopas.

Based on the image in the last 5 seconds of the video, it uses input based on the game graphics. The pixelated graphics of the game are used to represent symbols to specify what's what, like what's a brick block and what's the ground and what's a coin or what's a goomba or whatever. Also it appears to be using a reward system to motivate it to play the game. Inactivity causes a loss of 0.01 points per game tick, while doing stuff like jumping on a goomba or collecting coins grants positive points to counteract the passive loss per tick, while failure to complete the level causes -10 points, completely the level grants 100 points, etc. So if the AI's goal is "get as high a point total as possible" then it will want to get coins, jump on enemies, avoid death, and clear the level as fast as possible.

Neat.

[AlleeCat]

I wonder if the DM was setting up a tent while he was describing his bard. I mean, I like redheads, but that's absurd.

Well, I mean, generally if I'm not playing a redhead, I'm playing a lizard. But that's just me.

The DM is obviously being a total perv about his character, though. I don't mind sexy D&D as long as it's between consenting adults, but it seems as though the DM is the only one consenting here.

[BlackFlyme]

Dude's pretty much invited us all into his magical realm. Hopefully he doesn't manage to get everyone off of our current campaign and back into his.

Funny thing is that the games are hosted at Cleric's house. Investigator is Cleric's girlfriend, so I don't know how they felt about DM basically hitting on her via his DMPC.

Plus Cleric's parents are usually around while we playing, sitting and watching TV in the background or something. I wonder how they feel about him. I told my mom about some of DM's bull. She was dumbstruck.

If anyone wants to read the full greentext, I posted it here.

[Sergarr]

I wonder what sensory input that AI is getting, because it appears to be not enough to prevent Mario from repeatedly running into Koopas.

Based on the image in the last 5 seconds of the video, it uses input based on the game graphics. The pixelated graphics of the game are used to represent symbols to specify what's what, like what's a brick block and what's the ground and what's a coin or what's a goomba or whatever. Also it appears to be using a reward system to motivate it to play the game. Inactivity causes a loss of 0.01 points per game tick, while doing stuff like jumping on a goomba or collecting coins grants positive points to counteract the passive loss per tick, while failure to complete the level causes -10 points, completely the level grants 100 points, etc. So if the AI's goal is "get as high a point total as possible" then it will want to get coins, jump on enemies, avoid death, and clear the level as fast as possible.

Neat.

But wouldn't the constant loss of points cause the AI to constantly alternate it's behaviour, even if it's perfectly correct in-game, thereby ensuring that it would inevitably make a mistake in the future and putting an upper limit on how much it could adapt to Mario?

[Spehss _]

But wouldn't the constant loss of points cause the AI to constantly alternate it's behaviour, even if it's perfectly correct in-game, thereby ensuring that it would inevitably make a mistake in the future and putting an upper limit on how much it could adapt to Mario?

I dunno. ¯\_O_/¯

I watched a couple other videos of different AIs playing arcade games and the point reward system was used in them as well. So it seems to work for the goal of learning to play a game, at least.

AI plays Pacman. Reward score system can be seen in the bottom left corner of the game display. The score decreases by 1 constantly. Whenever the Pacman gets a power pellet it prioritizes eating ghosts over eating pellets so it can get more points while it can.
AI plays Centipede. Reward score system is seen displayed in the bottom right corner as "Reward: X points".

[BurnedToast]

That's the NEO DALTAN IMPERIAL, you earthbound swine!

You can't imagine my disappointment when that wouldn't fit in the naming box.

[Mech#4]

I remember one A.I created to play through Mario; it built up a knowledge on what to do based on several play throughs. Beginning with little steps left and right to moving forwards and eventually progressing through the level. The thing I found most interesting was the computer finding bugs and exploiting them to drastically increase its score. I also remember one where the A.I paused the game rather than lose because it meant its score didn't decrease from dying due to time out.

[BurnedToast]

I remember one A.I created to play through Mario; it built up a knowledge on what to do based on several play throughs. Beginning with little steps left and right to moving forwards and eventually progressing through the level. The thing I found most interesting was the computer finding bugs and exploiting them to drastically increase its score. I also remember one where the A.I paused the game rather than lose because it meant its score didn't decrease from dying due to time out.

I have to wonder if this is the AI actually learning the game, or if it's just "brute forcing" it.

If you made a brand new mario level, a human player who had learned the game would probably do pretty well at it, even on his first try. Would the computer have to start all over with the little left to right movements, or would it actually manage to finish the level the first try?

Brute forcing a game like mario brothers is interesting in it's own way, but if the computer was actually learning how to play the game in a more abstract way it's a lot more impressive.

[wierd]

The AIs from the MarioAI projects use heuristic sampling based on data presented on the screen, with some simple overriding objectives (Go as far right as possible, collect coins as most expedient, kill enemies as expedient.) 

Thus, they should be able to perform regardless of the level layout, as long as the layout is sensibly playable.

[Frumple]

So I'm apparently watching a playthrough of the (a?) 3DS Shantae game. Player gets to this couple, one of which has been petrified by a dread pirate's curse. Does stuff, gets the ancient spell that did the deed. Brings it to the unpetrified half the lovelorn pair.

Dude promptly reads off the spell, petrifying him, too. You'd think that'd be enough -- united in unending stone-encased horror -- but no. The combined weight of the two crashes through the ground, sending the pair plummeting into the depths. You follow after, and all that's left is a pile of rubble at the bottom of the drop.

It's, like. Holy shit. Game, you've been fairly lighthearted so far, and now you just brutally murdered two people. Not even haha comic silly "just kidding :V" murder, either. Stone cold dead. Not even enemies! Sympathetic characters, one of which just wanted to save his lifelong love from a terrible curse.

---

Let me also say that that game has one of the sickest bass lines I've heard in any game, which is just really gorram weird to see in something made for a handheld. It's got some pretty damned pumping tunes.

[Rolan7]

I remember one A.I created to play through Mario; it built up a knowledge on what to do based on several play throughs. Beginning with little steps left and right to moving forwards and eventually progressing through the level. The thing I found most interesting was the computer finding bugs and exploiting them to drastically increase its score. I also remember one where the A.I paused the game rather than lose because it meant its score didn't decrease from dying due to time out.

Sounds like this one: https://www.youtube.com/watch?v=xOCurBYI_gY
Gameplay starts a bit after 6 minutes.  The guy released some other videos about it too.  I love the tetris part's ending the best.
But, like BurnedToast suggested, this AI "brute-forces" in a way by calculating future states and choosing one.  Not sure how far ahead it looks, seems to be about 1-2 seconds.  I'm not even sure how much it learns, I think a lot of its behaviors (like trying to move right) were hardcoded.  Or maybe progress to the right was given value in its internal algorithm...  Same thing, I suppose.

[Sergarr]

The AIs from the MarioAI projects use heuristic sampling based on data presented on the screen, with some simple overriding objectives (Go as far right as possible, collect coins as most expedient, kill enemies as expedient.) 

Thus, they should be able to perform regardless of the level layout, as long as the layout is sensibly playable.

But sensibly playable layouts are boring! I would much more like to see AIs that can clear one of those insane ROMhacked levels, rather than see it beat something designed to be easy for little children.

[Spehss _]

The AIs from the MarioAI projects use heuristic sampling based on data presented on the screen, with some simple overriding objectives (Go as far right as possible, collect coins as most expedient, kill enemies as expedient.) 

Thus, they should be able to perform regardless of the level layout, as long as the layout is sensibly playable.

But sensibly playable layouts are boring! I would much more like to see AIs that can clear one of those insane ROMhacked levels, rather than see it beat something designed to be easy for little children.

This provides video footage of an AI sent to the Mario AI competition that uses A* search algorithm to "brute force" the game. This AI is later used to attempt to play through a level generated by a different AI, with this level designed to be "playable" but still with max difficulty that the generator AI can provide.