Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Pac-Man recreated with a GAN trained on 50k game episodes

blogs.nvidia.com

Post reply on HN
Print view View original

21–30 of 154 posts

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#21

Post by Ericson2314 »

ttul wrote: post #14
ekianjo wrote: post #13

> Trained on 50,000 episodes of the game Meanwhile an actual human will probably need only about 5 games to fully understand how the game works and reacts.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

Humans intelligence basically evolved (as far as we know) by retarding childhood development so we had neural plasticity longer to learn more. Crazy that we're giving our already-dumb AIs the extra challenge of learning things from nothing and doing so quickly.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#22

Post by amelius »

ttul wrote: post #14
ekianjo wrote: post #13

> Trained on 50,000 episodes of the game Meanwhile an actual human will probably need only about 5 games to fully understand how the game works and reacts.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#23

Post by cyberbanjo »

amelius wrote: post #22
ttul wrote: post #14

Earlier quoted context omitted.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

No, in humans the cycles involved in walking and 3d orienting carry over, so several thousand iterations of pre-game and a few of game.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#24

Post by ben_w »

itchyjunk wrote: post #8

Thought experiment: Is my brain also generating reality without a physics engine and is filling in details that it thinks should be there? If true, what are the advantages/disadvantages of it?

Absolutely.

A few times a year I might wake up in the middle of the night and — at first glance — perceive a goblin sitting on a nearby chair. As more of me wakes up, I realise it’s a pile of my clothes.

Now I’m learning German, I have the experience of listening to a conversation in German, thinking I’ve heard a particular word, asking the locals what it means, and finding neither of them noticed that word being used — but I genuinely hear the unspoken word, so I know a pre-conscious part of my mind is filtering all the German I hear through a vocabulary which is too small.

My mother has Alzheimer’s, and one of the bigger surprises was realising quite how many reality-models we have by way of her losing them at different times. When I was caring for her, star jumps terrified her, she forgot what windows look like at night and how to count past 5, and she lost both object permanence and the concept of left: https://kitsunesoftware.wordpress.com/2018/01/31/alzheimers/

One of the advantages of having a model of the world is it allows you to predict the future.

The only disadvantage I can think of is that it is comparable to branch prediction failure in a pipelined CPU — for example it can be very tempting to apply that to fellow minds, and when those minds don’t think like you, you can annoy them no end by telling them (implicitly or explicitly) that you know better than they do what they do and don’t want. Or you might become convinced that wearing lucky pants causes your sports team to win/your lottery numbers to come up. Or etc.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#25

Post by mcphage »

amelius wrote: post #22
ttul wrote: post #14

Earlier quoted context omitted.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

It might. My understanding is, it is common to take trained models from one problem and use parts of them on other problems in a similar domain to speed up training.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#26

Post by lonelappde »

pbhjpbhj wrote: post #17
ekianjo wrote: post #13

> Trained on 50,000 episodes of the game Meanwhile an actual human will probably need only about 5 games to fully understand how the game works and reacts.

I think a few more. Early games will establish things like "can ghosts go through each other", "if I hit the wall do I die" (like in Snake; though on Acorn Electron you could eat the wall sometimes!), "can ghosts go through the shortcut", "how long do pills last", do all pills last the same amount of time". Then later efforts would be needed to know things like "what algorithm do the ghosts follow", "does the same co…

Something being glossed over is the difference between being able to play the game and being able to implement the game.

Also, Nvidia's thing doesn't do the fruits and possibly not even eating ghosts or points and probably not the ghost algos.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#28

Post by saystupidthings »

amelius wrote: post #22
ttul wrote: post #14

Earlier quoted context omitted.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

I don't think that's what he is saying, but that is true. transfer learning for domain adaption is quite common, not sure how it works for behaviors, but if you were using say... a common controller for both the games then yes, I imagine it would take significantly less iterations to train.

Re: Pac-Man recreated with a GAN trained on 50k game episodes

#29

Post by lonelappde »

amelius wrote: post #22
ttul wrote: post #14

Earlier quoted context omitted.

Fundamentally, this is because we have already been trained on years of living in and reacting to the real world, which, to some abstract degree, the game mimicks. The Nvidia model starts with nothing.

So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

This called transfer learning and is an active area of research. It doesn't work nearly as well as it "should" if "AI" was really Artificial "Human Intelligence".
Post reply on HN