as a long time StarCraft II fan I am very happy, that I can now make this beautiful game the topic of my master thesis.
During my thesis I will, and to a certain degree already have, implement a number of different Reinforcement Learning algorithms that I will teach to play StarCraft II atleast in small combat scenarios.
For now I have 4 such scenarios that I would like to train my algorithms on.
And in order to get a good idea of how well the algorithms are doing I would like to compare them to human players of different skill groups.
For this reason I would appreciate it very much, if some of you could take the time and play these scenarios you can find here:
and record your scores(shown in the top left of the screen), to fill in here:
The 4 Scenarios in question are:
Find Ultralisk with Creep
This is more like a "Hello World" scenario and doesnt have much to do with actual StarCraft II. Kill as many 1hp Ultralisks within the time limit as possible. Creep instantly kills you if you step on it.
Blink Stalkers vs Roaches.
Standard Blink Stalker Micro. Points are awarded for each Roach killed and for every Stalker alive at the end of the scenario.
Reapers vs Zergling
Abuse cliffs in order to kill the group of Zerglings as fast as possible.
Points are awarded for each Zergling killed and each Reaper alive at the end.
Points are deducted for each 8 Seconds the Scenario is running.
Gateway Army vs Roach Ling
Use Forcefields, Guardian Shield, etc. to deal with the Zerg army.
Points are again awarded for each enemy killed and for each friendly unit alive at the end.
As there is some randomness involved it would be best, if you could do multiple tries of these scenarios and then give an average score, but I would be grateful for all data.
Additionally, if any of you have any ideas for other small, easy, but still interesting combat scenarios I could test my algorithms on I would appreciate that as well.
As soon as I have some solid data from my algorithms I will keep this post updated with scores and replays of my algorithms, and maybe even release my code and trained models should people be interested.