I'm sure it's interesting for those looking for incremental improvements in these sort of "reward box" type games where maximizing a set of attributes either directly counts as a win (i.e. not dying while reaching the end of a linear stage) or will inevitably lead to a win (StarCraft). There are a ton of games that, while not being arcade-y on the surface, have arcade mechanics at their core (like Doom), but I have yet to single AI that could play these sorts of games where there are multiple rewards that shift their priority constantly.
I did find one paper from many moons ago suggesting an RL AI that could play Doom, but all it could really do was shoot enemies on screen and pick up weapons, it couldn't actually finish a level (won't mention the fact it played on easy mode because I'm sure after enough generations it would become more skillful). Sure, learning the relationships between the various controls and the output on the screen is one thing, but I consider it a mostly solved problem. What would really interest me is an agent that can learn what "winning" is in the context of the game they're playing without knowing upfront. I'm fine if that requires training, but finding the win condition on their own always seemed to me as the obvious hard problem of these RL agents that papers always avoid tackling. On Sat, Jul 23, 2022, 8:43 PM <[email protected]> wrote: > > https://ai.googleblog.com/2022/07/training-generalist-agents-with-multi.html?m=1 > > They seem to be saying they are finding success by instead of training on > games using RL the original way, they are making it sort of watch poor/ > intermediate/ expert level gameplays and then just ask it to play as an > expert? > > So like I think this means (?) instead of ex. trying to move a little > space ship left/right at the bottom of a 2D screen, learning how to stay > not hit by bullets and hit alien ships say, it simply watches the videos of > us playing and uses that data to know what to do. It still needs the game > outcome scores and such then if so, so it knows which videos to think about > with if asked to play as an expert. So say a ship is coming close and a > bullet is shot at you, it does what seen, in the AI sort of way though of > course - it will probably choose to move over but also shoot back first (as > a rough idea). Using this big data they already have they seem to say, they > can skip training as if had no videos of people playing the game, because > we do have those videos. Which, would allow scaling better I guess they are > saying. Or help a ton. > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/T9d72fa28614a2681-M0c8123c96a232d930f5d4ffc> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T9d72fa28614a2681-M9e7657cca9ea38b00358180e Delivery options: https://agi.topicbox.com/groups/agi/subscription
