Deleted member 219079
D
Deleted member 219079
On Jan 24 the DeepMind team held a presentation for their Starcraft II AI "AlphaStar".
5 replays of AlphaStar going toe-to-toe with Team Liquid's TLO (PvP) were broadcast. In each of the matches TLO was pitted against a different agent, each with their own quirks. Next was 6-game showmatch against Liquid MaNa (PvP), comprising of 5 replays and 1 live game. Agents that were chosen to compete against MaNa ranked higher than those versus TLO in AlphaStar League, where ever-evolving agents battled each other for 14 days. The fittest of agents enlived 200 years of in-game time.
Some critique towards AlphaStar followed the event:
What I didn't expect was that cheesing agents made it far enough to be elected for games 1 and 5 against MaNa. Especially game 5 agent was quite the cheeser, you can see it casted by WinterStarcraft here.
What are your thoughts on AlphaStar? Are you curious to see more?
Edit: Find Artosis' analysis over the games here and an article released by the DeepMind team here.
5 replays of AlphaStar going toe-to-toe with Team Liquid's TLO (PvP) were broadcast. In each of the matches TLO was pitted against a different agent, each with their own quirks. Next was 6-game showmatch against Liquid MaNa (PvP), comprising of 5 replays and 1 live game. Agents that were chosen to compete against MaNa ranked higher than those versus TLO in AlphaStar League, where ever-evolving agents battled each other for 14 days. The fittest of agents enlived 200 years of in-game time.
Some critique towards AlphaStar followed the event:
- In the first 10 games the AI saw the whole map that led to it not having to position the camera to perceive game state and issue orders. In essence, anything not in fog of war was invariably observed by AlphaStar.
- While the average APM was told to be ~280, high amount of those actions are effective moves. Also the APM restrictions in place were too lenient; in game 4 versus MaNa, AlphaStar peaked at over 1000 EPM. For comparison, pro players rarely go above 300 EPM.
- Interacting through a custom interface, the AI doesn't have to box select units, allowing for unreproducible selections.
What I didn't expect was that cheesing agents made it far enough to be elected for games 1 and 5 against MaNa. Especially game 5 agent was quite the cheeser, you can see it casted by WinterStarcraft here.
What are your thoughts on AlphaStar? Are you curious to see more?
Edit: Find Artosis' analysis over the games here and an article released by the DeepMind team here.
Last edited by a moderator: