Discussion Topics

Purpose, Goals, and Obstacles

A1) In what ways is RTS AI research contributing to AI in general? Are we
    happy with this aspect of RTS AI research? How might we improve this
    aspect if not?

A2) What are the biggest obstacles toward achieving professional level RTS AI?
    How can we overcome them?

A3) Why do we see so few Zerg bots? Is StarCraft when bots play it imbalanced?

A4) Playing against bots is not the same as playing against humans (some
    strategies are better against humans but not bots and vice versa). Do we
    have to incorporate more machine vs. human games?

Tournament Format

B1) Do we have to limit the APM in tournaments to play a fair game against
    humans? Do we have to restrict the map view to sectors as well?

B2) By revealing the map pool that is used in tournaments in advance, on-line
    map analysis is mostly irrelevant and strategies can be tailored to maps
    manually. What are the obstacles to random map generations in StarCraft,
    which would make for an interesting thesis project?


C1) About benchmarking an RTS AI I was thinking that it will be great if we
    can define a list of specific RTS problems (like chess problems) to test
    how "mature" bots are.

Attracting Interest

D1) How can we lower the entry barrier to the RTS AI field? Right now it takes
    so long to get started with a basic bot that you need to be doing a whole
    PhD on the topic to even bother trying.

D2) What about the lack of new bots in the competition or the lack of
    bot-branches from previous bots entries? Maybe we need to improve the
    documentation of the bots.

Specific AI Techniques

E1) What I always see coming up when students work on problems like RTS is the
    hierarchical nature of strategy game AI. You have low-level tactics
    (operations), medium-level tactics (tactics), and high-level tactics
    (strategies). How to let AI systems learn in such a way that it performs
    well at all levels of tactics? I would like to discuss learning in
    strategic AI without handcoding a lot, and without cutting off areas of
    search from the outset.

E2) How about writing a simulator for abstracted RTS games and organize
    tournaments to improve low-level AI? Simulation based approaches will need
    such abstractions anyway.