Forums - Gaming Discussion - Live: Deepmind AI vs Tl's TLO Starcraft II

Blizzard and Deepmind tech have decided to show off the Deepmind AI going off against Team Liquid's TLO with a few games of Starcraft II:

https://www.youtube.com/watch?v=cUTMhmVh1qs

What's interesting about this so far, is that the AI itself seems pretty decent at micro managing it's own units, while keeping it's economy up constantly, while also improving it's strategies each time it clashes with TLO. 

From the Battle.net news reel:

When we last checked in with DeepMind, Oriol Vinyals stepped on to the BlizzCon 2018 stage to share the exciting progress their AI had made in StarCraft II. The AI, or agent, was able to perform basic macro focused strategies as well as defend against cheesy tactics like cannon rushes.

 

It’s only been a few months since BlizzCon but DeepMind is ready to share more information on their research. The StarCraft games have emerged as a "grand challenge" for the AI community as they're the perfect environment for benchmarking progress against problems such as planning, dealing with uncertainty and spatial reasoning.

Source: https://news.blizzard.com/en-us/starcraft2/22871520/deepmind-starcraft-ii-demonstration

Last edited by Chazore - on 24 January 2019

                                       

Around the Network

Hm.. if they needed to play Protoss v Protoss as the AI isn't trained enough in other match-ups yet, then they should've let it play against a Pro player that actually mains Protoss



Lafiel said:
Hm.. if they needed to play Protoss v Protoss as the AI isn't trained enough in other match-ups yet, then they should've let it play against a Pro player that actually mains Protoss

Afaik, they've been training the AI with neural networks (They call "Agents"), which have also played 5v5 matches, to which that data is then handed to the Alpha Star AI, which is technically like having 5 Alpha one AI's working as one unit at the same time.

The next upcoming match (within a few minutes) seems to be Alpha Star vs a pro Polish Protoss player. 

 

Edit: The AI just beat the Protoss player. 

Last edited by Chazore - on 24 January 2019

                                       

This is very interesting to see. AlphaStar is scary.



German YouTuber and Streamer:
StarCraft and Fallout 4 currently.

Me trying to write reviews:
Octopath Traveler

Amazon thread

Marth said:
This is very interesting to see. AlphaStar is scary.

It clearly doesn't care or have any sort of fear when it steps into someone's base, effectively stepping all over the human player economy. 

Usually with most built in game AI, the AI either gives up near the first/second signs of losing (even AoE AI is guilty of this), but with this Deep Mind AI, it just seems to keep going and going at it strong and almost precise. 



                                       

Around the Network
Chazore said:
Lafiel said:
Hm.. if they needed to play Protoss v Protoss as the AI isn't trained enough in other match-ups yet, then they should've let it play against a Pro player that actually mains Protoss

Afaik, they've been training the AI with neural networks (They call "Agents"), which have also played 5v5 matches, to which that data is then handed to the Alpha Star AI, which is technically like having 5 Alpha one AI's working as one unit at the same time.

The next upcoming match (within a few minutes) seems to be Alpha Star vs a pro Polish Protoss player. 

 

Edit: The AI just beat the Protoss player. 

I watched the matches they showed against TLO and they were rough from both sides, but the games against Mana (a top tier european protoss - not quite on the lvl of the korean protosses though) are much cleaner.



Lafiel said:

I watched the matches they showed against TLO and they were rough from both sides, but the games against Mana (a top tier european protoss - not quite on the lvl of the korean protosses though) are much cleaner.

Most definitely cleaner. Watching the AI losing that many Stalkers and having to deal with a stealthy counter was something that spiced things up.

 

I would love to see Deep Mind using their AI to train human players for championships (Mana even admits to learning some new things from his time against the AI, so they can help sometimes). 



                                       

Lafiel said:
Chazore said:

Afaik, they've been training the AI with neural networks (They call "Agents"), which have also played 5v5 matches, to which that data is then handed to the Alpha Star AI, which is technically like having 5 Alpha one AI's working as one unit at the same time.

The next upcoming match (within a few minutes) seems to be Alpha Star vs a pro Polish Protoss player. 

 

Edit: The AI just beat the Protoss player. 

I watched the matches they showed against TLO and they were rough from both sides, but the games against Mana (a top tier european protoss - not quite on the lvl of the korean protosses though) are much cleaner.

The Agents against MaNa had two more weeks of training time afaik. Which are another 200 years of playing against each other.

Though it went superhuman in Game 4. This level of Blink Stalker micro is impossible for a human.



German YouTuber and Streamer:
StarCraft and Fallout 4 currently.

Me trying to write reviews:
Octopath Traveler

Amazon thread

Marth said:
Lafiel said:

I watched the matches they showed against TLO and they were rough from both sides, but the games against Mana (a top tier european protoss - not quite on the lvl of the korean protosses though) are much cleaner.

The Agents against MaNa had two more weeks of training time afaik. Which are another 200 years of playing against each other.

Though it went superhuman in Game 4. This level of Blink Stalker micro is impossible for a human.

Yea, in Game 4 it finally showed the almost perfect unit control we feared was possible with super high multi-tasking.



Lafiel said:
Marth said:

The Agents against MaNa had two more weeks of training time afaik. Which are another 200 years of playing against each other.

Though it went superhuman in Game 4. This level of Blink Stalker micro is impossible for a human.

Yea, in Game 4 it finally showed the almost perfect unit control we feared was possible with super high multi-tasking.

Could we even hope to reach a level like that, if we were to say, "merge" with technology in the future?.

I think the thing that scares me the most, is that they had already brought both players into their home office prior to this recent demo reel, and both of them were beaten 5 times in a row, with zero wins against the AI. You would at least think that they would win one match. 

Last edited by Chazore - on 24 January 2019