3 ms·
1. HW resources: games run on extremely tight HW budgets. It's probably the most competitive, cut-through industry in terms of what you can afford to do with a
by Radim 10y ago
1. HW resources: games run on extremely tight HW budgets. It's probably the most competitive, cut-through industry in terms of what you can afford to do with a given CPU/memory/bandwidth per frame. While DL is still in the "let's throw a thousand GPUs at it, wait for a week, see what happens".
2. Quality control: DL is a blackbox. Bad for the playing experience for obvious reasons. And even if just using pre-trained models, the lack of control goes directly against what both the designers and players typically want (predictable fun).
3. Finally, DL hasn't made THAT much progress here, actually. There have been serious efforts to apply DL to modern games (e.g. DeepMind with Starcraft), but what you've probably seen is "just" DL applied to some finite-state, 100% deterministic, complete-information, zero-sum games. A far cry from complex dynamic environments.
Which is not to say it cannot be done. Just that the incentives (both business and technological) are aligned differently -- too little bang for a lot of buck.
- lloeki 10y ago> HW resources: games run on extremely tight HW budgets Come on, you could have decent AI back in the day that pinned me down and made me run for cover, then had me at that cling gling oh shit moment when a perfectly timed grenade lands at your feet just as you were starting to take a breath†. Today we're stuck with scripted sequences and playing whack-a-mole with sniper rifles. † only HL1, Halo and FEAR seemed to manage that.
- Radim 10y agoYou may be under appreciating how much effort goes into good old hacking in games, to make things look "natural" but with minimal HW footprint :) Cutting corners. Especially back in the day! It's more of an art. Applies to graphics, AI, audio...
- dkersten 10y agoQuality control: DL is a blackbox. I remember a discussion on gamedev.net once where exactly this was brought up. Its very hard or even impossible to tune an ANN to behave how the designers want it to. Simpler systems may not be as technically intelligent, but their behaviour can be tuned, encouraged, scripted etc to provide the experience that the designers wish. Beyond that, I remember they also talked about how it simply wasn't worth it because game AI isn't about being smart, its about being fun, interesting and "looking smart" (ie looking humanlike, human flaws and all). Machine learning is simply solving a different set of problems than what games have.