|
My predictions for the Student StarCraft AI tournament (SSCAI)
![[image loading]](https://i.imgur.com/JfaIDEm.png)
With the biggest StarCraft AI tournament now starting, it is time for some predictions on who will win. For those who don't know, since 2009 programmers have created AI to play StarCraft. Since 2017, StarCraft 2 has got itself an API for creating bots as well, but that one wasn't fully functional until later this year.
So the StarCraft AI tournaments with the most participators are still found in BW. And today the largest of those has started. You can watch it live on twitch:
https://www.twitch.tv/sscait
Since I mainly focused on SC2 this year and thus haven't updated my bot, I don't have a lot invested in the outcome. So this year would be a good time to start making predictions on who will win instead.
In my opinion, these are the bots to look out for this year:
SAIDA
![[image loading]](https://user-images.githubusercontent.com/34500497/47363581-16668200-d712-11e8-9f77-2b8fb0877e3e.jpg)
SAIDA is created by Samsung. It got 1st place at the previous big AI tournament "AIIDE", beating facebook who got 2nd place (Facebook isn't participating in SSCAI). A lot of hard coded rules, but they are adding more machine learning. This tournament they have a dedicated machine learning folder. So it will be interesting to see how much impact their new machine learning will have.
The previous version on the SSCAI ladder got 2nd place before the tournament started. So it will have a very good chance of winning this year's tournament. Even the new version seems vulnerable to some very specific rush builds though, so it is not 100%.
Locutus
One of the bots that is capable of defeating SAIDA. It got 4th place in the AIIDE tournament. With its variety of libraries and possible opening moves, it can catch opponents off guard. Build upon SteamHammer, which is a decent bot in its own right.
BananaBrain
Another bot capable of defeating SAIDA. Has seen a lot of updated for SSCAI. Didn't participate in AIIDE.
PurpleWave
A bot that got quite high in AI tournaments last year. Didn't participate in AIIDE, but got quite high in the SSCAI ladder before the tournament started. Doesn't seem to fare well vs SAIDA.
Iron bot
A bot that won the AIIDE in 2016. It got 6th place in the 2018 AIIDE. It struggles vs SAIDA, but is overall a decent bot. Solid mech play allows it to win slow and steadily.
McRave
Got 7th place in the AIIDE, but stands a decent chance vs SAIDA. Good at recognizing which opening the opponent uses, and reacting accordingly.
Marian Devecka
With its new update, it is trying to be the spiritual successor to the Berkeley Overmind. Good mutalisk control, but not enough to defeat SAIDA.
SteamHammer
SteamHammer got 8th place at the AIIDE, but Jay Scott has been adding a lot of new techniques for SSCAI. You can follow his progress on his blog:
http://satirist.org/ai/starcraft/blog/
He made some predictions about these bots himself. Overall, I agree with his sentiment, but I would say that Marian Devecka needs to play hydralisk games a lot better. Its over-reliance on mutalisks makes it more of a top 10 bot than a top 3 bot.
Overall predictions
To no surprise, I think that SAIDA will be in the finals. Due to the new changes in other bots, I'd predict that it will play vs one of {Locutus, BananaBrain}. Whether or not it will win that depends on if it can counter the new strategies that these two protoss bots use. So far, they seem to be able to defeat SAIDA, but if the new machine learning in SAIDA works well, it should adapt and win.
It will be interesting to see if machine learning will finally make a big difference in StarCraft AI tournaments, or if more hardcoded/scripted approaches will still remain dominant for a while.
|
This is amazing, I've been watching this for 30 minutes straight and it's so much fun lol.
I loved that one bot that only built marines non stop trying to put it into a single bunker, then when he hit critical mass, which was around 25 marines - or maybe more - he moved out. Only problem was there were 4 lurkers parked outside his choke. There was no detection, there were no medics, there was no micro.
And then the process of massing marines recommenced once again.
Ultimately the game dragged on a little too much because the Zerg waited to mass mutalisks and that took forever, but man that was fun.
|
On December 23 2018 04:31 Uldridge wrote:
This is amazing, I've been watching this for 30 minutes straight and it's so much fun lol.
I loved that one bot that only built marines non stop trying to put it into a single bunker, then when he hit critical mass, which was around 25 marines - or maybe more - he moved out. Only problem was there were 4 lurkers parked outside his choke. There was no detection, there were no medics, there was no micro.
And then the process of massing marines recommenced once again.
Ultimately the game dragged on a little too much because the Zerg waited to mass mutalisks and that took forever, but man that was fun.
Hopefully new players won't judge overall performance of the tournament and lose interest, just by observing few games which happened to be the weakest possible bots playing (ranked at bottom) those bots are long forgotten with latest updates like 4-6 years ago and even then they weren't anything special.
Wait for one of the top 10 or so bots to play to really see good bot playing.
To add upi think everybody understamete Proxybot which shows really fast progress for short time. I think it will be definitely top 16 if not top 10.
|
Hey I loved that stuff, so I'm not gonna lose interest in any way haha
|
I've been occasionally following Brood War bots for years, and I think it's really interesting that machine learning (neural networks, right?, or is it random forests) is finally being incorporated into some bots. And it's not dominant; instead it looks right now that if you work really hard with machine learning you might be incrementally better than hand-crafted bots.
What percent chance do you think SAIDA has to win SSCAI? Greater than or less than 50%?
I'm thinking of asking a question regarding SSCAI 2018 to a prediction website. Science and technology enthusiasts would make predictions on that question. (The website is called Metaculus.) Which question would you think is better?:
Will a Terran bot win the Student Starcraft AI Tournament?, or
Will the SAIDA Starcraft bot win the Student Starcraft AI Tournament?
or some other question?
|
On December 24 2018 06:20 datscilly wrote:I've been occasionally following Brood War bots for years, and I think it's really interesting that machine learning (neural networks, right?, or is it random forests) is finally being incorporated into some bots. And it's not dominant; instead it looks right now that if you work really hard with machine learning you might be incrementally better than hand-crafted bots. What percent chance do you think SAIDA has to win SSCAI? Greater than or less than 50%? I'm thinking of asking a question regarding SSCAI 2018 to a prediction website. Science and technology enthusiasts would make predictions on that question. (The website is called Metaculus.) Which question would you think is better?: Show nested quote +Will a Terran bot win the Student Starcraft AI Tournament?, or
Will the SAIDA Starcraft bot win the Student Starcraft AI Tournament?
or some other question?
SAIDA is looking good so far:
https://sscaitournament.com/index.php?action=scoresCompetitive
So I would give it more than 50%.
As for the prediction question:
"Will the SAIDA Starcraft bot win the Student Starcraft AI Tournament?"
Would be the best one to ask. I guess you can add some description describing what the SSCAI and specifically SAIDA is right? If so, add that.
|
On December 25 2018 12:03 LetaBot wrote:Show nested quote +On December 24 2018 06:20 datscilly wrote:I've been occasionally following Brood War bots for years, and I think it's really interesting that machine learning (neural networks, right?, or is it random forests) is finally being incorporated into some bots. And it's not dominant; instead it looks right now that if you work really hard with machine learning you might be incrementally better than hand-crafted bots. What percent chance do you think SAIDA has to win SSCAI? Greater than or less than 50%? I'm thinking of asking a question regarding SSCAI 2018 to a prediction website. Science and technology enthusiasts would make predictions on that question. (The website is called Metaculus.) Which question would you think is better?: Will a Terran bot win the Student Starcraft AI Tournament?, or
Will the SAIDA Starcraft bot win the Student Starcraft AI Tournament?
or some other question?
SAIDA is looking good so far: https://sscaitournament.com/index.php?action=scoresCompetitiveSo I would give it more than 50%. As for the prediction question: "Will the SAIDA Starcraft bot win the Student Starcraft AI Tournament?" Would be the best one to ask. I guess you can add some description describing what the SSCAI and specifically SAIDA is right? If so, add that.
Thanks, I wrote up and submitted the question and it was accepted.
|
|
|
|
|
|
EPT
