5 ms·

Have they though? I think I've run into realizations that deep learning is too expensive or not yet scalable enough to be useful in some applications, but like

Have they though? I think I've run into realizations that deep learning is too expensive or not yet scalable enough to be useful in some applications, but like other commenters on this thread. It can feel like a lot of the criticisms are grounded in some sort of feeling about what true intelligence should look like and an idea that number crunching can't possible be it

Retric 5y ago
It depends on what you mean by machine learning, but in practice stuff like video game enemy AI generally avoids deep learning for many reasons behind cost or scale. Time, tuning, complexity, and consistency for example are major issues in the real world. Sure the SCII AI eventually became very strong but it’s also somewhat immersion breaking to see all the odd stuff it ends up doing. And that’s with a well tuned game, in something under active development you end up with a host of other issues.
- space_fountain 5y ago
  Those are good points. Worth noting that video game AIs also often cheat. The goal for a video game ai is less to play the game well then to convince the player that they played the game well, but I think there definitely is something about believability or maybe human likeness that weirdly sometimes deep systems get very badly wrong
  - Jensson 5y ago
    Applying deep learning to computer game AI is just really hard. Not even a data center worth of compute and a big team of world leading scientists could create an AI able to beat the best players at games like starcraft or dota. It isn't reasonable to spend many times more more money making an AI for a game than you spent making the entire game, and the AI can't even run on any normal computer so you can't even use it at the end.
    - space_fountain 5y ago
      Yes though critically if you did want to actually write an AI capable of beating good players deep learning would probably be important. It has successfully been used to win at starcraft and dota now right?
      - Jensson 5y ago
        I agree, I think deep learning is necessary to create human class AI bots, but the current state of the art requires more compute than anyone can afford. Also applying that to computer strategy games is extremely hard to do, so you can't just sprinkle a bit on top and hope to make it better, it requires a lot of manpower from world leading experts to get results. This means that deep learning as of now isn't very useful for making strategy game AI's. > It has successfully been used to win at starcraft and dota now right? They didn't manage to beat the best in either game, but they did get close to top players in both games.
    - PeterisP 5y ago
      Wait, weren't both Starcraft and Dota getting very good AI results a few years ago by deepmind's alphastar and openAI respectively, IIRC they had some competitions with world champions? And I believe for the purpose of generalization and research they intentionally made the problem harder by using the visual render/representation instead of obvious technical simplifications of just taking the game state as any normal game bot would do. I mean, playing well is obviously not a goal for making the AI for a single-player mode or a bot for a multi-player game; the goal there is to play in an entertaining way, and the optimal system probably is one that is simple, easy to test, and is beatable in a manner that's fun for the players. Probably the only niche (very narrow!) where you would actually want near-optimal AI is for making a "sparring partner" for training e-sports players.
      - adgjlsfhk1 5y ago
        they got good results, but only in highly controlled, short duration testing. they're near the deep blue point, but still far from the era where humans are convincingly and resoundingly beaten.

Retric 5y ago

It depends on what you mean by machine learning, but in practice stuff like video game enemy AI generally avoids deep learning for many reasons behind cost or scale. Time, tuning, complexity, and consistency for example are major issues in the real world. Sure the SCII AI eventually became very strong but it’s also somewhat immersion breaking to see all the odd stuff it ends up doing. And that’s with a well tuned game, in something under active development you end up with a host of other issues.

space_fountain 5y ago
Those are good points. Worth noting that video game AIs also often cheat. The goal for a video game ai is less to play the game well then to convince the player that they played the game well, but I think there definitely is something about believability or maybe human likeness that weirdly sometimes deep systems get very badly wrong
- Jensson 5y ago
  Applying deep learning to computer game AI is just really hard. Not even a data center worth of compute and a big team of world leading scientists could create an AI able to beat the best players at games like starcraft or dota. It isn't reasonable to spend many times more more money making an AI for a game than you spent making the entire game, and the AI can't even run on any normal computer so you can't even use it at the end.
  - space_fountain 5y ago
    Yes though critically if you did want to actually write an AI capable of beating good players deep learning would probably be important. It has successfully been used to win at starcraft and dota now right?
    - Jensson 5y ago
      I agree, I think deep learning is necessary to create human class AI bots, but the current state of the art requires more compute than anyone can afford. Also applying that to computer strategy games is extremely hard to do, so you can't just sprinkle a bit on top and hope to make it better, it requires a lot of manpower from world leading experts to get results. This means that deep learning as of now isn't very useful for making strategy game AI's. > It has successfully been used to win at starcraft and dota now right? They didn't manage to beat the best in either game, but they did get close to top players in both games.
  - PeterisP 5y ago
    Wait, weren't both Starcraft and Dota getting very good AI results a few years ago by deepmind's alphastar and openAI respectively, IIRC they had some competitions with world champions? And I believe for the purpose of generalization and research they intentionally made the problem harder by using the visual render/representation instead of obvious technical simplifications of just taking the game state as any normal game bot would do. I mean, playing well is obviously not a goal for making the AI for a single-player mode or a bot for a multi-player game; the goal there is to play in an entertaining way, and the optimal system probably is one that is simple, easy to test, and is beatable in a manner that's fun for the players. Probably the only niche (very narrow!) where you would actually want near-optimal AI is for making a "sparring partner" for training e-sports players.
    - adgjlsfhk1 5y ago
      they got good results, but only in highly controlled, short duration testing. they're near the deep blue point, but still far from the era where humans are convincingly and resoundingly beaten.