Hacker News Reader

Comment by WhitneyLand | Hacker News Reader

What indeterminate games can AI can beat humans at?

Backgammon - AI usually beats the best humans

Ms. Pac Man - AI loses to almost any human

AI tends to perform poorly for many real-time games because they constitute many, many more "moves" than a game like backgammon. A backgammon game will typically take < 100 moves. For a game like Ms. Pac Man the AI will have to make that many decisions in a timespan of about four seconds. When you train a reinforcement learning algorithm, the AI has to update the parameters that were used to make several thousand decisions, and the only information it has to go on is its score at the end of the game. (This is known as the credit assignment problem --- of all the decisions you made in the game, only a few probably had a substantial impact on the outcome. How do you decide which ones were the important ones?)

A game like go is hard because there are many possible moves every turn. But the games themselves are not very many turns, so it is easier to find the important moves, because in some sense all the moves were more important. A game like Ms. Pac Man is hard because, while you only have 5 moves each turn, the games themselves are many, many turns, and it is very difficult for the NNs to learn these long-term dependencies.

Since a child could easily outperform the best AI here, doesn't it imply we need a different approach?

Otherwise we need 4 orders of magnitude more compute, and it's just hard to believe their is no better way.

No, just means that someone needs to bother to make an AI for Ms. Pacman.

Handling random ghost movements is only trivially confounding for AI learning. Low value, low yield.

It's true that someone could pretty easily hand-craft an AI that plays Ms. Pac-Man pretty well. But that's not really the point. Because after all that work, you still only end up with a computer program that can play Ms. Pac-Man. It won't play Pong, it won't play Doom, it won't even be able to play regular Pac-Man! The goal is to build a general system that can learn how to play specific games on its own. So, there is potentially high value to developing a Ms. Pac-Man AI, but only if it's not really a Ms. Pac-Man AI, but actually a more general AI, for which playing Ms. Pac-Man is a specific instance.

I agree with the parent comment that some sort of new approach is probably needed. The main issue with reinforcement learning is just that the algorithm gets too little feedback. This ends up discarding an enormous amount of data on each run (or, more accurately, learning from that data extremely inefficiently). There's a huge amount of data present, but the learning from that data has to be unsupervised. A child can do it easily --- they only have to play Ms. Pac-Man once or twice to get the hang of it. Computers currently cannot. Getting them to do that will definitely require something different, but no one is quite sure what.

>someone could pretty easily hand-craft an AI that plays Ms. Pac-Man pretty well

Not true. Many people have also tried handcrafting as well and only scored around 30,000 which is a low score.

How much more complex is Ms. Pac-Man when compared to Atari 2600 games? Deep Minds DQN did master (as in better then any human can play them) a whole bunch (49) of 2600 games a while back

https://deepmind.com/research/dqn/

edit: Checking wikipedia looks like it isn't playing all games it can play at superhuman levels

> For most games (Space Invaders, Ms Pacman, Q*Bert for example), DeepMind plays below the current World Record

https://en.wikipedia.org/wiki/DeepMind#Deep_reinforcement_le...

DM did terrible at Ms. Pac Man, let alone below the current world record is an understatement.

DM scored less than 30,000. Human record is over 900,000.

I think you are underestimating the difficulty. It's not like DM was the first to try.

Multiple papers have been written on failed attempts. Dozens of other developers have tried.

It's an open problem.

It's a boring tautology. Arcade games require "clever" (read: not generalizable) solutions to optimize for speed in decision making.

"Learning" AI requires the broadest feasible search space for any solution and successful designs require clusters of machines. Clusters of machines have network latency and coordination overhead.

This is the same latency vs throughput design decision that every system must make, and its not impressive when a throughout oriented system struggles with being latency sensitive.

It may be boring because you don't realize it's not a tautology.

You are wrong to separate video games from generalized AI. If you read the paper I posted above it explains part of this active area of research,

Conventional AI can clean up at Pacman if it can do tree search on the game state, source: CS188 out of Berkeley.

http://ai.berkeley.edu/lecture_videos.html

Pac Man is easy. Ms. Pacman has been tried and failed by many people who know CS188 material.

Do these attempts make use of the documented idioscynracies of Ms Pacman?

http://gameinternals.com/post/2072558330/understanding-pac-m...

Note that most of recent work in reinforcement learning focus on creating a model which works on all or most of Atari games without using any game-specific knowledge. If researchers were to focus on only Ms. Pac Man, and ignore the "no game-specific tuning" rule, then Ms. Pac Man, and most Atari games would be solved to near perfection. The challenge that the Atari suite provides is not their individual difficulty, but the difficulty imposed by their diversity, i.e. finding a general algorithm which works well on all Atari games is hard.

I don't know why so many people are suggesting this is easy.

Many other attempts has been made targeted specifically against Ms. Pac Man and none have come anywhere near beating an average human player.

What makes Ms. Pac Man so hard for AI?

I'm not sure it's hard in general but deepmind's program didn't do well because it couldn't plan ahead https://www.technologyreview.com/s/535446/googles-ai-masters...

It's also inferring the game state from looking at the screen rather than being spoonfed the data.

One could conceivably perform a depth-limited search on the actual game state if it were available and then use an AlphaGo-like DNN to predict what a deeper search would find, no?

Dunno really. A human would watch the ghosts behaviour and guess their likely future behaviours based on that. I'm not sure if the software gets that, as it were, or how you'd tweak it to do so.

Its indeterminate because the ghosts turn in random directions with no preserve patterns. Notice Go, Chess, etc all are determinant.

This paper gives an overview of the background: http://www.cse.unsw.edu.au/~mit/Papers/AAAI10a.pdf

Reply on news.ycombinator.com