5 ms·
Its seems that the current progress in AI pioneered by DeepMind is predicated on being able to measure progress by using a pre-defined set of rules (i.e. the ga
by 10dpd 11y ago
Its seems that the current progress in AI pioneered by DeepMind is predicated on being able to measure progress by using a pre-defined set of rules (i.e. the game currently being played).
How will this work in a MineCraft simulation where there are no rules? Where will the feedback loop come from, e.g. a subjective evaluation by "humans"?
- BoudewijnE 11y agoAs in the limitations u have in minecraft = rules.
- TuringTest 11y agoYou still can create a set of goals within the game that count as "success" and others as "failure" to guide the reinforcement; and have the learning process try to figure out everything else in Minecraft in order to achieve those goals.
- zamalek 11y agoIndeed, a GA would work quite nicely here - but I guess you'd need a cluster of machines simulating (not playing) Minecraft for this to be a viable experiment.
- DanBC 11y agoMinecraft survival has player health and player armour strength. It also has a hunger bar. And the player has an inventory of items collected. And then there are a bunch of statistics. http://minecraft.gamepedia.com/Statistics http://minecraft.gamepedia.com/Statistics I agree this would make a player that has an unusual play style (never making a base, for example), but it might make for some fun Let's Play videos. And increasing YouTube subscriber count, and video view count, and video thumbs up, while minimising video thumbs down would be an interesting experiment. ("I welcome our AI Clickbaiting overlords")
- zdkl 11y agoI would follow that channel and its devblogs assiduously
- waterlesscloud 11y agoI'm much more interested in an AI that optimizes audience stats than one that optimizes in-game stats.
- soared 11y agoI predict a twitch stream of an AI playing various games. Watching it fail and succeed on various games and situations would be very interesting. Maybe a couple devs to do a voice-over.
- TOMDM 11y agoIf you haven't seen it already, you might be interested in subreddit simulator https://www.reddit.com/r/subredditsimulator https://www.reddit.com/r/subredditsimulator which is pretty much that but for upvotes on reddit. It's at most kind of amusing at times for now
- Houshalter 11y agoThis is something I discussed with some fellow AI enthusiasts 2 years ago. We tried to come up with some "reward function" that would encourage very general behavior and reward more generally intelligent AIs. The idea I had was to reward collecting blocks of different types, with exponentially decreasing reward. So finding a block it's never found before is 100 points, the second it finds is 50 points, and so on. It encourages exploration and touching on as many mechanics of the game as possible. The important part is that the AI learns on its own though, not programmed to do specific tasks by a human. However it should perhaps be given recipes and some other meta information, because discovering those by chance is highly improbable.