Doom, episode 1 level 1... Visual Doom AI’s CIG competition pits bots only that play against each other using nothing but visual references rather than internal data. (image via id Software, via Visdoom)
Academic groups and institutions have been pitting their respective AI platforms against humans in various games for a number of decades now in an effort to train their systems to learn using any number of discipline methods. The latest of which was demonstrated by Google’s recent Go match between their AI AlphaGo and world champion Lee Sedol (Lee resigned in 4 out of 5 games) back a few weeks ago.
Playing board games is one against human opponents is one thing, playing a legendary FPS like the original Doom against other AI contenders is another, which is what researchers from ViZDoom are looking to accomplish with their Visual Doom competition for this year’s CIG (Computational Intelligence and Games) conference.
The competition revolves around AI players only pitted against each other in a death match where the contestants have access to one weapon only, in this case the rocket launcher but are able to grab health and ammo resupply crates as well. The contest is set with two maps on two contest tracks with the first being a ‘limited’ contest with a known map that programmers can familiarize their AIs on. The second will revolve around three unknown maps and be a full-on deathmatch, albeit with the single weapon and pickups only.
Each AI player will have their own PC from which to play and will battle each other for 10-minutes on a single map with the limited track being played for 12 rounds and the second track running for 4-minutes. Like real-world human death matches, the winner will be chosen based upon who has the highest frag count.
Programmers (and anyone for that matter) have full access to the ViZDoom environment created for the researchers visual learning project, which can be found on GitHub and includes the necessary APIs as well as programming and trouble shooting examples for Python, C++ and Java- the only languages that are being allowed.
Unlike some of the other AI learning platforms which rely extensively on ‘machine learning’ and ‘tree searching’ techniques to learn to play (Google’s AlphaGo uses Monte Carlo tree search). These platforms make use of technical data including map layout, object placement and the position of other players.
ViZDoom utilizes visual input only, meaning the AI can only extrapalate information from what is projected on the monitor- in essence, what it see’s is what it gets and makes decisions based on that information in much the same fashion as a human. Those interested in joinging the competition must submit their addmissions either by March 31st for early testing matches or by August 15 for final submissions. No prizes have been annponced as of yet, however ViZDoom is currently looking for sponsers.
Have a story tip? Message me at: