eSports / videos

[StarCraft II] DeepMind demonstration of AI capabilities
I love these demonstrations of Artificial Intelligence development in games, because this is the best way to show us the progress of capabilities of AIs and the way they "think" in comparison to us humans. Last week we had a showcase of DeepMinds newest AI presented in the DeepMind Headquarters in London. For those that don't know much about StarCraft no worries since commentators Daniel "Artosis" Stemkoski and Kevin "RottterdaM" van der Kooi done a good job of explaining everything that's going on in the presentation followed by some DeepMind team members giving more clarification about workings of their newly developed AI AlphaStar. AI was tested versus 2 pro players from Team Liquid, Dario "TLO" Wünsch, and Grzegorz "MaNa" Komincz. For those interested in this watch the VOD of the stream, it's a lengthy video so I will leave some timestamps below: Timestamps: (note: they didn't show all the games on the stream, to see all the replays visit this link.)29:38 - Beginning of the stream and introducing TLO43:58 - Game 1 TLO vs AlphaStar (after that game is a brief introduction on how AlhpaStar "sees" the game)58:38 - Game 3 TLO vs AlphaStar (after that game DM team present a bit how AlphaStar trains itself)1:26:40 - Introducing MaNa1:32:25 - Game 1 MaNa vs AlphaStar1:42:01 - Shows how AlphaStar evaluates the game state1:46:05 - Game 3 MaNa vs AlphaStar2:00:02 - Game 4 MaNa vs AlphaStar2:21:23 - DeepMind video about the development of AlphaStar2:31:22 - Live Exhibition Match, MaNa vs AlphaStar [SPOILER ALERT] This part is for those that choose not to watch the video: Although for now, AlphaStar is limited to play on one map (Catalyst) and in a match-up Protoss vs Protoss only, the performance is really impressive. To test its true capabilities and to get the best possible feedback about this project DM team decided to invite some pro players to see how it compares to some of the top players in StarCraft II. In the first 5 game match versus TLO, a strong player, but not a main in Protoss race got whipped 5-0! AlphaStars superiority looked really convincing to the DM team and they wanted to make it a bigger challenge and invited TLO teammate from Liquid, MaNA. MaNas main race was indeed Protoss so that should show how strong this AI really is in this kind of match-up. Yet again, AlphaStar whipped the pro player 5-0! At this time, it seemed that the AI has really some potential to become unbeatable even by the strongest players. Before those games, players were informed on how AlphaStar is working and how it perceives the game and makes decisions and after the match both players agreed that they were at a major disadvantage because of the AI had the ability to "see" the whole map all at once and was aware of all the units at all times, basically like it is zoomed out to see the entire map, which is impossible for the players, not only because of human capabilities but also because the game didn't allow that. So after that feedback and before the stream above, DM team decided to make a different iteration of Alphastar that has this same restriction of only seeing what is currently on the screen, in the way normal gameplay would be. That iteration was again tested, but this time live on the stream versus MaNa. The early game seemed like it's going to go in the same fashion as previous matches, but MaNa figured he will "stretch" the attention to a couple of places on the map in the same time and it was proven to be effective at confusing the AI making it go back and forth with its units thus buying time to grow his army strong enough and finally push for the victory! Conclusion It seems that DeepMind is developing their AIs rather quickly, that's why they are the world leaders in the AI technologies. They showed that when it is presented with all information needed about a closed system (games) in a perfect information environment it can be a matter of days or even hours that the AI will master the game and develop strategies beyond human comprehension, as AlphaZero did in Chess. But in those imperfect information environments like in StarCraft II where the attention is limited to the camera's perspective and with the mechanic "fog of war", AI struggles to make the best decisions. I think that kind of shows how our intuition helps us make decisions on "gut feeling" and previous experience when we don't have all the necessary information but still we can predict most situations that can occur during the game. What do you guys think about this? Please let me know in the comments. For more information on this project visit DeepMinds website. If you want to hear MaNas personal experience while being involved in this project and his analysis of the games he played, watch this video: Thank you for reading! , z3ll