4 ms·

> Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations

> Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations it was building) but simply out-microed its opponent, it just had better mechanics. Protoss Alphastar was fantastic. It was definitely not just out-microing opponents (after later refinements Alphastar's micro limitations were very human-like), but really had an understanding of build orders, positioning, choke points and the like. Here's a great game it played against Serral (one of if not the top Zerg player in the world) where it holds fantastically against Serral's aggression and is able to use the layout of its own base to pick apart what looked like a winning position for Serral: https://www.youtube.com/watch?v=C6qmPNyKRGw https://www.youtube.com/watch?v=C6qmPNyKRGw

tornato7 6y ago
As an avid StarCraft player I still think AlphaStar is one of the most impressive AI programs ever made. While it didn't get everything quite right, it exhibited planning and creativity in it's builds, and didn't just do the same exact thing every time; it scouted and reacted to the opponent's moves with professional-level proficiency.
- CyanBird 6y ago
  It didn't react it had pre-planned build orders That's something which was mentioned when it came to deep mind team talking of all the "different agents" they had trained
  - dwohnitmok 6y ago
    As far as I understand it did actually react. AlphaStar's training process consisted of two phases: an imitation-of-human-players learning phase that created a base set of "templates" and then self-play among agents seeded from those templates. When creating those templates, they basically ended up with "baked-in" build orders. But self-play among agents ended up creating a variety of reactive strategies on top of those build orders, so that agents would modify them on-the-fly as the game situation changed. At a high level this is basically how humans use build orders as well.
Retric 6y ago
Later refinements where still quite superhuman, just in terms of consistency not abusing stalker micro. For example it doesn’t need control groups, that’s a clear sign it’s not sticking close to human limitations.
- onethought 6y ago
  Yeah the alpha star team didn’t really emphasise it. It’s APMs were human like in count, but they had basically 0 wasted actions, they were all meaningful. Lots of APM are used by repeat keys or mass clicking.
  - dwohnitmok 6y ago
    > Lots of APM are used by repeat keys or mass clicking. Both of those (and more) are filtered by EAPM measures (effective APM) and even on those metrics Alphastar still had a lower EAPM count than most pros.
    - Retric 6y ago
      EAPM still includes a lot of pointless clicking by pros. Just look at periods of the game when there is nothing productive to do. By comparison a throttled Aplphastar has no need for any pointless actions and can get timings inhumanly precise which games EAPM in the other direction.
- dwohnitmok 6y ago
  Kind of. The only reason it didn't use control groups is because it had "soft" control groups of its own (IIRC there were limits on how granular of actions it could take as well as how many distinct move commands it could issue). I mean looking at that Serral game I don't see any superhuman micro on the part of Alphastar. Just fantastic exploitation of base layout and steadily pressing every strategic weakness Serral made in his attack. If anything on the consistency front, Alphastar's main advantages came from superhuman macro (effectively never forgetting about making sure workers are busy), but the edge there vs human pros is pretty slight.

tornato7 6y ago

As an avid StarCraft player I still think AlphaStar is one of the most impressive AI programs ever made. While it didn't get everything quite right, it exhibited planning and creativity in it's builds, and didn't just do the same exact thing every time; it scouted and reacted to the opponent's moves with professional-level proficiency.

CyanBird 6y ago
It didn't react it had pre-planned build orders That's something which was mentioned when it came to deep mind team talking of all the "different agents" they had trained
- dwohnitmok 6y ago
  As far as I understand it did actually react. AlphaStar's training process consisted of two phases: an imitation-of-human-players learning phase that created a base set of "templates" and then self-play among agents seeded from those templates. When creating those templates, they basically ended up with "baked-in" build orders. But self-play among agents ended up creating a variety of reactive strategies on top of those build orders, so that agents would modify them on-the-fly as the game situation changed. At a high level this is basically how humans use build orders as well.

Retric 6y ago

Later refinements where still quite superhuman, just in terms of consistency not abusing stalker micro. For example it doesn’t need control groups, that’s a clear sign it’s not sticking close to human limitations.

onethought 6y ago
Yeah the alpha star team didn’t really emphasise it. It’s APMs were human like in count, but they had basically 0 wasted actions, they were all meaningful. Lots of APM are used by repeat keys or mass clicking.
- dwohnitmok 6y ago
  > Lots of APM are used by repeat keys or mass clicking. Both of those (and more) are filtered by EAPM measures (effective APM) and even on those metrics Alphastar still had a lower EAPM count than most pros.
  - Retric 6y ago
    EAPM still includes a lot of pointless clicking by pros. Just look at periods of the game when there is nothing productive to do. By comparison a throttled Aplphastar has no need for any pointless actions and can get timings inhumanly precise which games EAPM in the other direction.
dwohnitmok 6y ago
Kind of. The only reason it didn't use control groups is because it had "soft" control groups of its own (IIRC there were limits on how granular of actions it could take as well as how many distinct move commands it could issue). I mean looking at that Serral game I don't see any superhuman micro on the part of Alphastar. Just fantastic exploitation of base layout and steadily pressing every strategic weakness Serral made in his attack. If anything on the consistency front, Alphastar's main advantages came from superhuman macro (effectively never forgetting about making sure workers are busy), but the edge there vs human pros is pretty slight.