3 ms·
slightly related to this, I read somewhere recently that deep reinforcement learning doesn't really work. Has there been any progress on that front since?
by frequentnapper 7y ago
slightly related to this, I read somewhere recently that deep reinforcement learning doesn't really work. Has there been any progress on that front since?
- FartyMcFarter 7y agoAlphaGo seemed to work.
- ghaff 7y agoSo a digital perfect information game. It's pretty well-accepted that this is a great use case for RL. The question is how much broader the scope is.
- dqpb 7y agoIt worked for Dota.
- gautamcgoel 7y agoI recently attended a panel discussion on machine learning at the ITA workshop. I asked the panel (several distinguished ML profs among them), which idea was most overrated in ML? The answer: deep RL.
- MasterScrat 7y agoThere's been great success to play video games, initially Atari games and more recently Dota and Starcraft II. Real world achievements are still lacking. Do check the recent advances from Covariant in factories though: https://www.wired.com/story/ai-helps-warehouse-bots-pick-new-skills/ https://www.wired.com/story/ai-helps-warehouse-bots-pick-new...
- sgt101 7y agoWell - all I can say is that it works for me! More seriously, I think that it can be challenging to define the "game" or simulation that a DRL system (or any RL) system depends on. Clearly for things like go the game is preset, for the real world, perhaps less so. People have used it vs physical systems like robot hands though. Also, one of the charms of DRL is the cheating that the agents discover, for example catapulting each other over barriers or blocking doors. But in real world scenarios scams that break the rules of the game (imposed by law or physics) are useless, so you have to rewrite your simulators to remove them, and start again.