Agent Demo List #1
Description
This issue is a collection of various interesting agent demonstrations trained by ACE, it will be updated continually.
SMAC
smac.mp4
We carefully analyze the behavior of transferred agents. As shown in the above video, ACE perfectly transfers the station position, careful positioning, alternating fire, and focusing fire of 5m vs 6m to 8m vs 9m and then achieves 40% win rate in 8m vs 9m. However, the optimal choice of focusing fire in 8m vs 9m is different from that in 5m vs 6m, which is largely responsible for ACE not achieving a 100% win rate on the 8m vs 9m map. We describe the transferring experiments in detail below
-
ACE perfectly transfers the station position of 5m vs 6m to 8m vs 9m, making the agents stand up and down in a line.
-
ACE perfectly transfers the operation of careful positioning and alternating fire of 5m vs 6m to 8m vs 9m. Each ally has very little health but kills all enemies.
-
The optimal choice in 5m vs 6m is all allies focusing fire on one enemy. In comparison, the optimal choice in 8m vs 9m is all allies focusing fire on two enemies, since 5 allies can kill an enemy instantly, and having 8 agents attacking one agent at the same time is wasteful.
GRF
- academy_3_vs_1_with_keeper
academy_3_vs_1_with_keeper.mp4
- academy_counterattack_hard