(D/C rank) man vs machine TeamLeague - Page 8

Grats!

grats as well! Score is 3-1 for in favor of Team Sauerbraten. Last Match is gonna be Artanis vs FlashFTW and will be scheduled soon

Artanis fault

This is a highly philosophical stance that has major backing in the science community, or at least had for decades. The idea that AI can exist in a vacuum, as in without humans telling it what to do, is a highly debated subject. The core of the theory is - if it is humans that create AI, would the AI ever outgrow the parents? If we are AI's creators, then how can the AI ever exceed the human capacity? It's an interesting thought piece and although logic may dictate that AI can never surpass its maker, the top scientists of the previous generation have already been proven wrong on multiple counts as to the capability of AI. We are in the infant stages of it. It is too early to tell what AI can or cannot do.

More on the subject, after hearing the casts and playing against multiple AI myself, I have to say that the most immediate goal is to create algorithms that mimic situational awareness. LetaBot's creator kept saying that "this algorithm needs to be improved" or "this is on the to-do list." In my eyes, if you could create an AI that would analyze replays from an FPV, then it would be able to learn how to counter every strategy in the game. What I mean by this is, when you run a replay in BW, you are basically forcing two computers to recreate the actions committed by either player. If LetaBot could assume the role of Flash when watching replays, I think that it could possily not only gather information as to what build orders are good, what reactions are necessary to other builds (since it would have access to the actions of the opponent), it would also learn timings and movement. Being an AI and not a human, it could analyze these things at x16 or even from parsing the data contained within a replay, meaning that its learning potential is only limited by the strength of its learning algorithm and the replays fed to it. If one AI is fed Flash replays, another Jaedongs, and another Bisu's, and they were forced to simulate round-robin games for a couple of weeks non-stop, I am sure that even further progress will emerge. Of course, this is all easier said than done. However, I think that the philosophical approach to this issue which you used to argue that machines have no desire to win/no emotion is not the machine's disadvantage in this scenario. Their lack of pure creativity will still exist. In other words...

What I see now is that programmers are manually inputting building placements, build orders, adaptive strategies, etc. If this process could be automated, then an AI which would be much more capable of beating a C level player consistently will emerge quite rapidly given the vast amount of data (replays) and processing power (modern computers) available. Humans would be forced into a scenario where they know their mechanics cannot compete with the computer's, so they must present a scenario that the AI is unfamiliar with, knowing full well that after presenting the scenario once, the AI will learn of its existence and then it would try all of its existing builds in order to beat it. If adaptability is on a higher level than we currently expect, it may even blend strategies and ideas. If pure information gathering is on an extreme level and it was fed all currently existing XvY replays, surely at some point it will find a build that beats your "new" build, even though (theoretically) we know that no genuinely new ideas exist and that more likely than not, any build you try has won and lost 10 times in the history of Brood War. In short, an AI of that level would only benefit from every loss whereas players like mishin on ICCup prove that humans have a far lesser capacity for such growth.

That in my opinion is the future of AI in a competitive environment such as this. The Go bot has shown that it is possible. BW is a much more open battle field with much more resulting variance on input and therefore output, so it will take a few more years of development from dedicated programmers such as LetaBot in order for us to reach that point. However, all predictions for the development of AI have been beaten in other areas already, so maybe mine is subject to the same exponential error but it is what it is. All I can say is I am excited for the future, and that a bot that can beat Flash will be a reality, especially once his new ideas run out and he starts repeating builds from the 20 years of SC:BW.