printz Posted March 11, 2012 Here are a couple of Ultimate Doom vanilla demos recorded using my port-in-progress AutoDoom. For the most part in these demos, the player is moving automatically, except that I command it where to go, by pointing in the automap. Pathfinding and shooting monsters is done by the computer, as it can be well seen in the demos. First demo is E2M3 on HNTR, going until the exit. It's so simple it can be solved... Second demo is E1M1 UV, with 100% kills and secrets, also until the exit (you can verify the statistics by using PrBoom+ or Eternity and their HUD). The bot doesn't push the exit switch in any of them, because I haven't implemented the logic yet :)autodemo.zip 0 Share this post Link to post
dew Posted March 11, 2012 oh wow... that's crazy! also don't think i haven't noticed you are trying to obsolete us and put machines in our working places! we will go on a strike if further provoked! 0 Share this post Link to post
j4rio Posted March 11, 2012 If popularity of doom was comparable to some MMO, I'd bet somebody would come up with this sort of thing sooner or later. You simply go sleep and wake up in the morning with a bunch of computer generated demos on your HD that will be undistinguishable from a non-TAS demo because, well, it would be tweaked the way that one wouldn't be able to tell if it's TAS. I doubt it's worth the trouble though. 0 Share this post Link to post
printz Posted March 11, 2012 j4rio said:You simply go sleep and wake up in the morning Just keep in mind this isn't autonomous yet, it was actually responding to "attack-move" commands I was giving in the Automap :) 0 Share this post Link to post
abyrvalg Posted March 12, 2012 So, is there any (self-)education the bot can do? Or are you using a strict set of rules for the bot to follow, and he always performs exactly the same if you give him the same commands? Any way, this is really impressive and inspiring to see, stoked to see more of this stuff in action :) I was going to start studying neural nets to try applying them to doom, by training a net so that it can consistently run ep1 uv-max (it's ok if it would perform much slower than human). This is a huge goal and it's just a hobby (i've never studied it in university, although i could), so i'm waiting until summer to have more time for this project. It's really doubtful this thing would be able to beat any existing TAS records, but one useful application is to quickly test questionable places, i.e. cutting corners in nomo-runs. For reference, see nomo-built runs for map01: it's hard to decide where to run in the beginning - Sylvain Chabert runs to the western wall to wallrun it, others just go straight to the eastern corner to cut it, same with the final wallrun in the brown room. It'd be great if some heuristics could quickly provide additional information about such simple things and give TASer an idea of what should be tested and what shouldn't. Right now i'm writing a simple program which randomly generates a starting "population" of demos (or you can set them manually) and uses a simple genetic algorithm or particle swarm optimization (look up wikipedia if you're not familiar with these, they're pretty simple) to try and come up with a better lmp. This thing will probably suck, as it can only write GF50+SL50 tics and the only variable at each tic is a turning angle, but i'm still going to try it :p It's basically just a dumb, guided semibruteforce, but it worked well on simple models in my tests, almost always improving my own solution (here's the pics, the goal was to get in the green cross in a predefined number of "tics". The body moves with acceleration, but there was a typo in the law of motion, not a big deal though. Max speed is limited: http://dl.dropbox.com/u/17623935/3circles_touch1.jpg and http://dl.dropbox.com/u/17623935/3circles1rect_002.jpg) 0 Share this post Link to post