|
Poland3748 Posts
On January 25 2019 07:06 Kafka777 wrote:
This was interesting. Nevertheless right now AI is playing a different game. It has much higher effective APM, by far superior micro management and superior vision, it will not be distracted in terms of sight. That is why it won games.
The strategies presented by AI were rather poor and reactions questionable. I would say that AI has still a long way to go before it beats any pro players on equal terms, as successfull AI must prove itself beat players mainly by adopting superior strategies.
It's hard to say what is equal terms in terms of video game. One could argue that equal terms is when ai will control physical keyboard,. Mouse and get screen data from the display. But that's adds a lot of problems on top of in game decisions. The alphastar decisions were fine because it executed them with insane micro all the time. But the last game shown that alpha star is really playing best on what it seems in very recent period of time and the repetitive immortal drops has broken it. In that respect it is very alien, and I believe top players could adapt to its style and play around it.
|
One thing people also need to remember is that the blizzard measurement of EAPM is not actually all effective APM. Rather it is Blizzard EAPM = APM - "the most obvious spam clicks".
Alot of the time pro players will do slightly subpar clicks or they will compensate with higher APM and do extra clicks because they know their mouse accuracy isn't 100% accurate (very relevant when you do splitting vs AOE).
So a computer that can average 350 apm + spike is 1500 is absolutely gonna stomp human pro players as long as it's general strategic knowledge and decision making is "not terrible".
I would be more interested in seeing it capped at a mechanical level where it comes to mid-high GM players and then see whether it strategically could win games against top pro players.
For that to be the case, it would probably need to be capped at avearge of around 150 apm with spikes no higher than 500.
|
I think unless it's a robot sitting at a computer operating a keyboard and mouse, people are always going to be able to take shots at an AI Starcraft player. Unlike chess and go, the interface is a huge part of Starcraft.
Unfortunately, I'm guessing it's not beneficial for DeepMind to try and go that far, since their interest is purely in developing AI and neural networks. I'm sure they'll make a really good AI that stomps any human player, but it will always do it from inside the machine. Then, they'll move on.
Maybe someday a robotics team will take an AlphaStar brain and put it into a robot, who knows haha
That said, I wasn't expecting this iteration of AlphaStar to be anywhere near this good, so it was cool to watch.
|
On January 25 2019 16:49 Popkiller wrote:
I think unless it's a robot sitting at a computer operating a keyboard and mouse, people are always going to be able to take shots at an AI Starcraft player. Unlike chess and go, the interface is a huge part of Starcraft.
Unfortunately, I'm guessing it's not beneficial for DeepMind to try and go that far, since their interest is purely in developing AI and neural networks. I'm sure they'll make a really good AI that stomps any human player, but it will always do it from inside the machine. Then, they'll move on.
Maybe someday a robotics team will take an AlphaStar brain and put it into a robot, who knows haha
That said, I wasn't expecting this iteration of AlphaStar to be anywhere near this good, so it was cool to watch.
Even if there was a Terminator sitting in the chair they'd complain that steel wrists don't hurt. Artificial intelligence is called that for a reason.
Some people just like to whine. They whined when AlphaStar won, they whined when it lost, and they'll whine the next time it plays, no matter what accommodations Deepmind grants. I look forward to hearing their excuses when it wins again.
|
former deepmind machine learning scientist (and former sc2 player) here. i worked on a separate (unannounced) project but collaborated with a lot of the people who worked on alphastar, so i wanted to share some thoughts from an ML and SC perspective. these opinions are completely my own and don't reflect the positions of deepmind or the company i'm currently at.
ML:
- this is a pretty remarkable achievement in terms of demonstrating imitation learning and RL as a means to learn long term strategies.
- most of the algorithms (inverse planning, imitation learning) used are not new, but have never been applied in such a way and at this scale. a lot of crazy engineering went into this that most people don't think of
- there are still notable failure cases which is why there's still more to do to beat top pros in any matchup, but mostly just more training. this could theoretically generalize to any RTS game.
- this is closer to real life than any other existing environment ever created.
SC:
- as a former sc2 player myself, sure i'll admit i wasnt that impressed for reasons others have stated including:
1. alphastar had some physical advantages a human doesnt have
2. alphastar made some blatant "errors" from a gameplay perspective (not getting blink, unable to handle warp prism harass), the latter which mana exploited for his only win
3. i never really saw alphastar scout, then react to an opponent. it always seemed like it had a strategy at the start, then executed it.
- sc2 is a game where macro effects determine the outcome of a game - major battle losses usually. this means it's really hard to know if alphastar had a superior strategy or if it outmacro-ed the humans. and that was demonstrated tenfold by the fact that it had useless workers, 10 some observers, useless units, etc etc. and won those games convincingly.
at the end of the day, this is not to appeal to sc2 fans, just as AlphaGo wasnt meant for go fans (it's just a great side effect). it's to use these games as a testbed for real world situations - planning and decision making in partial information environments with exponentially large action spaces.
|
On January 25 2019 16:53 pvsnp wrote:
I look forward to hearing their excuses when it wins again.
everyone expects AI to win over time. that's what it does, it brute forces games in a way that not only can a human not do, but that humans collectively, as an entity, cannot do. winning is inevitable. but it is an empty exercise right now when the AI is winning by producing mass stalkers into a hard counter and then winning because it can blink micro in a way that a human being cannot. that's simply pointless.
|
Russian Federation40190 Posts
To be honest i am not going to mention any excuse but just state that it would be more interesting if it was a ZvZ, though i suspect Deepmind ZvZs never go past baneling nest.
|
Czech Republic12129 Posts
On January 25 2019 17:09 negativedge wrote:Show nested quote +On January 25 2019 16:53 pvsnp wrote:
I look forward to hearing their excuses when it wins again. everyone expects AI to win over time. that's what it does, it brute forces games in a way that not only can a human not do, but that humans collectively, as an entity, cannot do. winning is inevitable. but it is an empty exercise right now when the AI is winning by producing mass stalkers into a hard counter and then winning because it can blink micro in a way that a human being cannot. that's simply pointless.
Well, this AI didn't brute force it though. I mean AI with bruteforcing wouldn't kill its own units via big boom balls of doom against TLO 
|
brute force learning, not playing in an individual game
|
On January 25 2019 16:53 pvsnp wrote:Show nested quote +On January 25 2019 16:49 Popkiller wrote:
I think unless it's a robot sitting at a computer operating a keyboard and mouse, people are always going to be able to take shots at an AI Starcraft player. Unlike chess and go, the interface is a huge part of Starcraft.
Unfortunately, I'm guessing it's not beneficial for DeepMind to try and go that far, since their interest is purely in developing AI and neural networks. I'm sure they'll make a really good AI that stomps any human player, but it will always do it from inside the machine. Then, they'll move on.
Maybe someday a robotics team will take an AlphaStar brain and put it into a robot, who knows haha
That said, I wasn't expecting this iteration of AlphaStar to be anywhere near this good, so it was cool to watch. Even if there was a Terminator sitting in the chair they'd complain that steel wrists don't hurt. Artificial intelligence is called that for a reason. Some people just like to whine. They whined when AlphaStar won, they whined when it lost, and they'll whine the next time it plays, no matter what accommodations Deepmind grants. I look forward to hearing their excuses when it wins again.
If they didn't promote this as a fair match between pro and AI most ppl wouldn't have problem with it. This is PR based on at the very least a biased approach of what's going on.
|
In my opinion the truest level playing field is not a crippled bot vs a human playing with keyboard and mouse, but a bot vs a human whose brain is directly hooked to SC2, so that the human also gains the advantages that the AI had. Imagine if you also had access to perfect clicks and stalker micro. It's fun to think about-- I think the AI, at least the AI we have at today, would never beat a human hooked to SC2, ever.
|
brute force learning, not playing in an individual game
The term "brute force" has a specific definition in computer science. Generally speaking, it means to compute all of the possibilities and then select the best one. However, this is not feasible in Starcraft for a number of reasons. It wasn't even feasible with the game of go, and Starcraft is a lot more complex. AlphaGo and AlphaStar are both capable of quickly recognizing which possibilities are not worth exploring, but without having to actually evaluate the outcome of such actions. This is why some people describe their mode of operation as more intuition-based. The AI doesn't have to think about what would happen if it made obviously terrible decisions, like move commanding back and forth in front of the enemy's army.
If they didn't promote this as a fair match between pro and AI most ppl wouldn't have problem with it. This is PR based on at the very least biased approach of what's going on.
There is no way to spin this as simple PR. Writing an AI that can play RTS at a human level is a world-first achievement in computer science.
EDIT: What OpenAI did with Dota was a PR stunt. As far as I'm concerned, as long as the AI is playing real Starcraft and not some limited version, then it certainly qualifies as a fair evaluation of its strength.
|
On January 25 2019 17:44 imre wrote:Show nested quote +On January 25 2019 16:53 pvsnp wrote:On January 25 2019 16:49 Popkiller wrote:
I think unless it's a robot sitting at a computer operating a keyboard and mouse, people are always going to be able to take shots at an AI Starcraft player. Unlike chess and go, the interface is a huge part of Starcraft.
Unfortunately, I'm guessing it's not beneficial for DeepMind to try and go that far, since their interest is purely in developing AI and neural networks. I'm sure they'll make a really good AI that stomps any human player, but it will always do it from inside the machine. Then, they'll move on.
Maybe someday a robotics team will take an AlphaStar brain and put it into a robot, who knows haha
That said, I wasn't expecting this iteration of AlphaStar to be anywhere near this good, so it was cool to watch. Even if there was a Terminator sitting in the chair they'd complain that steel wrists don't hurt. Artificial intelligence is called that for a reason. Some people just like to whine. They whined when AlphaStar won, they whined when it lost, and they'll whine the next time it plays, no matter what accommodations Deepmind grants. I look forward to hearing their excuses when it wins again. If they didn't promote this as a fair match between pro and AI most ppl wouldn't have problem with it. This is PR based on at the very least a biased approach of what's going on.
"Fair" is one of those words that everyone has a different definition for. Machines aren't people. Neural networks don't have a whole lot in common with neurons besides the name. Could you please point to the cat in this picture?
![[image loading]](https://savan77.github.io/blog/images/im2.png)
But you are right about one thing. This (meaning the showmatches and such) is a PR stunt. The real success is in the major technical progress they've made. Deepmind doesn't really care about Starcraft except as a vehicle for improving our understanding of AI. The visibility and media attention is just a nice bonus.
To put this in perspective, Google is over a billion in the red because of Deepmind. That money is not being spent to win games of Starcraft. Fairly or otherwise.
|
|
|
Demanding robot arms is a bit silly. At some point it's like asking your mother, who is about to visit the market, to bring back a bouquet of flowers, including roses, tulips and the rare Edelweiss flower growing only at high altitude. It transforms the task from buying something at the market to becoming a mountaineer. Robot arms are a separate field of AI (literally the field of robotics), whereas Deepmind seeks to develop machine learning tools for planning sequential tasks in noisy environments with incomplete information and large action spaces, or whatever is the proper terminology. In that respect the throttling of APM and the added limitation to camera movement are besides the point, only a concession to the domain.
If you want AI to compete in fairer circumstances, such that its APM is hardcapped and noise and drag is introduced to its cursor control, then you have to pressure not Deepmind, but Blizzard. The latter is the only one with an actual investment in StarCraft. For instance, if they intend to use AlphaStar as a tool for balance testing, then it has to mimic human physical constraints to be useful; and if they intend to play more show matches and challenge top players, it has to be able to win convincingly by superior decision making, not act as if it was barely evolved beyond the meanest micro bot.
|
Poland3748 Posts
One disappointment for me was that AlphaStar really excelled with micro which is kind of mini-game where you kind of have full knowledge or at least information you have is not that imperfect.
In terms of long term planning with huge amount of imperfect information AlphaStar was really unimpressive - it was that it didn't matter that much as long as it could compensate with insane micro and good decent tactics.
It actually bogs the question how well DeepMind's approach would scale for longer games where imperfect information has even bigger impact on the outcome.
|
I bet Deepmind playing zerg vs terran or protoss it would be alot weaker, especially zvt.
Its hard to play safe against every possible openings terran can do against zerg. Zerg units also have a lot less micro potential. Sometimes as zerg you have to take calculated risks and drone a bit harder than should be safe.
Im a bit disapointed we only saw pvp... PVP is really simple and comes down mostly to execution and micro.
Would love to see it play ZvT where every game is a totally different opening from terran, when it would be getting attacked and harassed non stop all game.
We saw that it struggled once harassed with warp prism... and they had blink stalkers. I cant help but feel deepmind wouldnt stand a chance in ZvT currently.
I feel like PVP and ZVZ would be the easiest matchups for deepmind to dominate in. Because both those matchups are pretty straight up and all about micro and massing lots of the same unit.
|
What strikes me and people seem to completely forget about it, is that for their 5 best agents, they claimed that none was THE best. This means that each can loose to another one. So of course, as it was pointed out, if you study closely all 5 agents, learn their tendancies, then scouting should be enough to adapt and win.
This is also a very good news for us players, as it means that all strategies are probably viable, at least as a counter to another one.
Pro players would be comparable to 3 to 5 agents, with different BO planed. For example, one BO micro oriented with an blink all in, another one with a double WP harass, and another one with a fast carrier.
For now, AI agents still don't really know how to adapt and shift their strategies to match the one of the opponent. That's why the AI lost to MaNa. He adapted to the AI by harassing, and the AI was clearly confused.
Also claiming this is just a micro-bot is far from understanding how deepmind AIs work. They learn from scratch. I would have actually loved to see some kind of progression, see how the game was playing after 10 years, 50 years, 100 years. The only progression we saw was that one week training between TLO's game and MaNa, which was already an impressive improvement in the tactical parts and strategically (1 base carrier against TLO, mass phoenix against Mana)
the game vs TLO also showed that the AI learned to defend (probably not in the most efficient way but still) canon rushed. It probably means that other agents out of the top 5 were canon rushers.
Anyway, I'm quite impressed by the AI progress, from the beacon search to this
|
On January 25 2019 18:19 LDaVinci wrote:
What strikes me and people seem to completely forget about it, is that for their 5 best agents, they claimed that none was THE best. This means that each can loose to another one. So of course, as it was pointed out, if you study closely all 5 agents, learn their tendancies, then scouting should be enough to adapt and win.
This is also a very good news for us players, as it means that all strategies are probably viable, at least as a counter to another one.
Pro players would be comparable to 3 to 5 agents, with different BO planed. For example, one BO micro oriented with an blink all in, another one with a double WP harass, and another one with a fast carrier.
For now, AI agents still don't really know how to adapt and shift their strategies to match the one of the opponent. That's why the AI lost to MaNa. He adapted to the AI by harassing, and the AI was clearly confused.
Also claiming this is just a micro-bot is far from understanding how deepmind AIs work. They learn from scratch. I would have actually loved to see some kind of progression, see how the game was playing after 10 years, 50 years, 100 years. The only progression we saw was that one week training between TLO's game and MaNa, which was already an impressive improvement in the tactical parts and strategically (1 base carrier against TLO, mass phoenix against Mana)
the game vs TLO also showed that the AI learned to defend (probably not in the most efficient way but still) canon rushed. It probably means that other agents out of the top 5 were canon rushers.
Anyway, I'm quite impressed by the AI progress, from the beacon search to this
If a professional player elects for a different strategy for each game of the series, then he is still recognizably the same player, with the same patterns and habits in his control, such that you can predict his reaction to feints, charges, diversions. Furthermore, although a player can deliberate in opening choice, they will tend to fall back to their general style in the mid- and late game.
But suppose you play a game versus AlphaStar, you notice not only some higher level decisions such as a tendency to opt for blink stalker builds, but you also pick up on some habits such as bad scouting, inability to defend against harassment, ineffective wall-offs. And now you're considering your approach for next game. It's obviously perfectly proper for the agent to rotate between strategies in order to avoid being figured out, therefore you realize you can't blindly counter the previous blink stalker build. But you also can't count on its weaknesses in defense, because you're playing a different agent. You don't know if it will have good blink micro, because it's a different agent. etc. In its most extreme form, any type of decision which involves interaction with your opponent will be based on quicksand, because you're playing a completely new opponent every time with no history, no information about it. Whatever this is, it's not standard match conditions. It's more like playing five random ladder games versus barcodes, with the opponents being picked out of time and space, e.g. one will be a 2016 player from Korea, the other a 2018 player from Europe, the other a diamond level player that just does cannon rushes.
In short, it's not predictable, unless all the agents tend to converge to a similar style, which is speculative, but if that's the case then I think that humans could adjust and develop anti-AI strategies, similar to how computers could be defeated in chess for years despite superior calculation ability. Obviously AI's will win in the end, but it's an open question whether that's a week from now or five years from now. We don't even know if Deepmind will stay with the project long enough to thoroughly trash all human opposition. afaik there has been BW AI research for around 10 years without threatening pro gamers.
|
On January 25 2019 09:28 Garrl wrote:
I do have to wonder if the way it plays is informed by its' mechanical ability - if we gave it infinite APM and and a client with enough framerate to support it, would it just do totally fundamentally unsound plays because it can get away with it?
Yes if you train it against human opponents, not quite as it is trained today against other agents. It would do plays that might be fundamentally unsound if made by a human, but that would still be positively selected against opponents with the same APM.
On January 25 2019 13:11 ThunderJunk wrote:
That was amazing. Can't wait to integrate the Deepmind probe saturation into my pvp
If it is learnt (and not just a side effect of "building probes is good"), we have to assume the continued probe production had a positive effect on the agent's winrate during training.
Could mean that other agents on average do a better job harassing the economy than what human opponents did, so that the continued probe production that would have been necessary in face of a better harassment leads to oversaturation against humans.
|
|
|
|
|
|
EPT
