|
On January 25 2019 09:40 TheDougler wrote:
You don't know that it was the the camera change that actually was the determining factor here. It could be that Mana had a better idea of what he was up against. It could be that the warp prism threw off the AI's gameplan (I think it's this one). It could be that this AI isn't quite as good as other AIs.
Mana said the strats the AI came up with threw him off. Specifically walking up the ramp in game 1. It's something a pro would never do so he felt safe up until the point AS forced his way up the ramp. Likewise, Mana said in game 5 that AS's strat was so weird he had no idea what was going on. It caught him completely off guard.
If pros had time to study the agent's games, I think it would be much easier to beat them right now. The agents start with pro games to study, so it's only fair that humans get to study their games.
Obviously there's a lot to do on DeepMind's side before they reach the ultimate test. Obviously getting rid of raw interface and switching to the camera like the show match is a good first start. Playing non-mirror matchups is another, as well as on different maps.
The final I would say is to play only one agent. Every game used a different agent. It's akin to playing different players. TLO didn't know this when he was playing and played his matches as if it was the same agent and thus tried strats to counter what he just saw in the previous game, which of course didn't work. Playing against a single agent would be quite interesting.
|
I thought this was really cool to watch... would love to see more!
I loved seeing the strange things the computer did, like go mass stalker and over saturate the bases.
Things we would think of as non-optimal, and then wonder, "was the other thing really the most optimal way to play?"
|
On January 25 2019 11:48 counting wrote:Show nested quote +On January 25 2019 11:20 xuanzue wrote:On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither You need to download their custom map and put it in the Maps folder first, as described in their download page deepmind.comTo load the replays: Install StarCraft II. It is free to play and runs on Windows and Mac. Create the StarCraft Maps directory: Windows: C:\Program Files (x86)\StarCraft II\Maps Mac: /Applications/StarCraft II/Maps Download the map. Move it into the Maps directory. Download the replays above, and open them as usual.
the map links directs to a moma page, something only google employes can access.
|
Absolutely incredible work. Also quite scary how fast AI is advancing.
|
On January 25 2019 12:01 imp42 wrote:Show nested quote +On January 25 2019 11:28 No_Roo wrote:On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
You've got it exactly backwards. The people critical of this result are interested in the computer playing by the same set of rules, right now it's playing with extra rules that benefit it, which are not available to humans. Allowing it to see without the restrictions of a camera is a special rule that benefits the computer. Allowing it to have direct API access is a special rule that favors the computer. Starcraft is a game designed and balanced around physical limitations of the input & output devices used to play it. If you're not using those devices, or at least realistically observing their limitations, you're not playing starcraft. If you're impressed with the results, then hey, great. But it's not fair to say critics are the one demanding special privileges here. Maybe take an additional step back and notice that the game itself is already extremely unfairly biased towards humans (see https://www.teamliquid.net/blogs/518977-towards-a-good-sc-bot-p5-hi-2-2, section 1). From this point of view, we already play with a lot of rules that benefit us humans. Regarding the zoomed out camera view, I agree. That's clearly an unfair advantage for the AI. However, first results seem to show that having to control the camera just slows down the learning a bit and does not represent a significant obstacle. Direct API access is a different story. It would be relatively easy to engineer a robotic arm that physically uses a keyboard. It's just not valuable in terms of advancing the field of AI.
A lot of time, we forget how impressive our human vision and instant decision action learning are. From an engineering point of view, human vision and object detection mechanism is so well tuned and calibrated by eons of evolution that it can identify objects and evaluate them subconsciously and then reconstruct the world around us in our head from moment to moment and make assumptions almost right the first time most of the time given just a few examples.
All the above abilities are actually quite far off from realization in current engineering capability. That is someone might said it is easy to just install a robot arm, and a pair of camera as eyes for an AI system to achieve end-to-end learning (that is simulate a human constrain), which is completely unrealistic at the current moment. The reason for invoking API as interface, is more or less due to it is very difficult for pure vision object detection to do it efficiently and correctly, as well as the control for output with a robot arm require so many parameters to just create one action alone will need to be learned and tuned with infinite possibility. So it is the engineering limitation make this necessary. (imagine how badly an AI system will function, where 5% of the input is incorrect, and half of the output actions are not what it is intended to do, it will be impossible to train it no matter what, and image recognition will also reduce the reaction time, not to mention the training time. If raw image input to objectives/objects is easy, we will already have the reinforcement learning for playing any game already).
The use of full map view is also an engineering challenge related to the I/O limitation. Most people don't realize how amazing their "imagination" or "reconstruction" of the field of view really is. For an raw image input to jump all over the map and able to picture them together as a whole and find temporal correlations, and they changed and being covered by units/buildings all the time, is very very difficult. I am amazed to how well the limited view version actually function for the live match. That along is an achievement. And as the learning curve showed it took quite a while almost 2 to 3 days to overcome and realize the raw input of a map is actually a map and make coherence output is astonishing. if you put these to the time scale 7 days is like 200 years, it actually spend most time like decades of repeat to learn these which is actually not very smart but more brute force. I'd imagine it will have hard time introducing a new map which might take learning from the start to get this realization again (I believe it memorize landmark patterns in order to navigate and tell which part of the map it is looking at instead of generalized to any type of "virtual view" construction), and I believe it is also the reason why it's performance for microing from multiple angles are less impressive.
|
On January 25 2019 12:14 vesicular wrote:Show nested quote +On January 25 2019 09:40 TheDougler wrote:
You don't know that it was the the camera change that actually was the determining factor here. It could be that Mana had a better idea of what he was up against. It could be that the warp prism threw off the AI's gameplan (I think it's this one). It could be that this AI isn't quite as good as other AIs. [...] The final I would say is to play only one agent. Every game used a different agent. It's akin to playing different players. TLO didn't know this when he was playing and played his matches as if it was the same agent and thus tried strats to counter what he just saw in the previous game, which of course didn't work. Playing against a single agent would be quite interesting.
A misconception IMO. There is no conceptual difference between "one agent" and "multiple agents", because you can simply combine x agents into one composite agent (which is exactly what they did).
Compare it to Innovation switching up his macro game with a 3-rax proxy cheese. It's not akin to playing different players, but the same player choosing a different game plan before he starts the game.
The concept of a composite agent gets interesting when you add a super-agent to it, responsible for picking a sub-agent to play a specific game in a boX match. I would imagine the super-agent would then be trained similar to a Texas Hold'em agent and converge to game-theoretical optima for cheese / standard ratio etc.
|
T.O.P.
Hong Kong4685 Posts
On January 25 2019 12:33 xuanzue wrote:Show nested quote +On January 25 2019 11:48 counting wrote:On January 25 2019 11:20 xuanzue wrote:On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither You need to download their custom map and put it in the Maps folder first, as described in their download page deepmind.comTo load the replays: Install StarCraft II. It is free to play and runs on Windows and Mac. Create the StarCraft Maps directory: Windows: C:\Program Files (x86)\StarCraft II\Maps Mac: /Applications/StarCraft II/Maps Download the map. Move it into the Maps directory. Download the replays above, and open them as usual. the map links directs to a moma page, something only google employes can access.
I uploaded the map here
https://drive.google.com/file/d/1axKe_Q5qekI-8a6Skva5pWi5Hv_GMb-2/view
Follow the instructions above to get the replays working
|
On January 25 2019 12:33 xuanzue wrote:Show nested quote +On January 25 2019 11:48 counting wrote:On January 25 2019 11:20 xuanzue wrote:On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither You need to download their custom map and put it in the Maps folder first, as described in their download page deepmind.comTo load the replays: Install StarCraft II. It is free to play and runs on Windows and Mac. Create the StarCraft Maps directory: Windows: C:\Program Files (x86)\StarCraft II\Maps Mac: /Applications/StarCraft II/Maps Download the map. Move it into the Maps directory. Download the replays above, and open them as usual. the map links directs to a moma page, something only google employes can access.
I think too many people downloading at the same time, causing some issues. This is the backup I downloaded before and works.
CatalystLE.SC2Map
|
On January 25 2019 12:45 imp42 wrote:Show nested quote +On January 25 2019 12:14 vesicular wrote:On January 25 2019 09:40 TheDougler wrote:
You don't know that it was the the camera change that actually was the determining factor here. It could be that Mana had a better idea of what he was up against. It could be that the warp prism threw off the AI's gameplan (I think it's this one). It could be that this AI isn't quite as good as other AIs. [...] The final I would say is to play only one agent. Every game used a different agent. It's akin to playing different players. TLO didn't know this when he was playing and played his matches as if it was the same agent and thus tried strats to counter what he just saw in the previous game, which of course didn't work. Playing against a single agent would be quite interesting. A misconception IMO. There is no conceptual difference between "one agent" and "multiple agents", because you can simply combine x agents into one composite agent (which is exactly what they did). Compare it to Innovation switching up his macro game with a 3-rax proxy cheese. It's not akin to playing different players, but the same player choosing a different game plan before he starts the game. The concept of a composite agent gets interesting when you add a super-agent to it, responsible for picking a sub-agent to play a specific game in a boX match. I would imagine the super-agent would then be trained similar to a Texas Hold'em agent and converge to game-theoretical optima for cheese / standard ratio etc.
This actually has a technical term in machine learning community called ensemble learning. But I don't think it is that easy to implement as of yet. And for efficiency sake the single agent is actually very different from a group of agents which will absolutely require quite a bit of parallel processing to achieve (it is not as simple as installing more GPU can solve). And indeed these agents choose to represent the group of all agents in the AlphaStar league will be those encounter many different strategies and still win for the most part overall. It actually is a very difficult problem to introduce "novelty" and still able to adapt mid-game. The current system is simply not having any learning capability on the fly (within one game, in machine learning term, it is a system with offline learning, instead of active/online learning which is much much more difficult).
|
On January 25 2019 12:48 T.O.P. wrote:Show nested quote +On January 25 2019 12:33 xuanzue wrote:On January 25 2019 11:48 counting wrote:On January 25 2019 11:20 xuanzue wrote:On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither You need to download their custom map and put it in the Maps folder first, as described in their download page deepmind.comTo load the replays: Install StarCraft II. It is free to play and runs on Windows and Mac. Create the StarCraft Maps directory: Windows: C:\Program Files (x86)\StarCraft II\Maps Mac: /Applications/StarCraft II/Maps Download the map. Move it into the Maps directory. Download the replays above, and open them as usual. the map links directs to a moma page, something only google employes can access. I uploaded the map here https://drive.google.com/file/d/1axKe_Q5qekI-8a6Skva5pWi5Hv_GMb-2/viewFollow the instructions above to get the replays working
Thank you!
Also the instructions need more accuracy, because not always everyone will install the game in program files.
|
On January 25 2019 12:42 counting wrote:Show nested quote +On January 25 2019 12:01 imp42 wrote:On January 25 2019 11:28 No_Roo wrote:On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
You've got it exactly backwards. The people critical of this result are interested in the computer playing by the same set of rules, right now it's playing with extra rules that benefit it, which are not available to humans. Allowing it to see without the restrictions of a camera is a special rule that benefits the computer. Allowing it to have direct API access is a special rule that favors the computer. Starcraft is a game designed and balanced around physical limitations of the input & output devices used to play it. If you're not using those devices, or at least realistically observing their limitations, you're not playing starcraft. If you're impressed with the results, then hey, great. But it's not fair to say critics are the one demanding special privileges here. Maybe take an additional step back and notice that the game itself is already extremely unfairly biased towards humans (see https://www.teamliquid.net/blogs/518977-towards-a-good-sc-bot-p5-hi-2-2, section 1). From this point of view, we already play with a lot of rules that benefit us humans. Regarding the zoomed out camera view, I agree. That's clearly an unfair advantage for the AI. However, first results seem to show that having to control the camera just slows down the learning a bit and does not represent a significant obstacle. Direct API access is a different story. It would be relatively easy to engineer a robotic arm that physically uses a keyboard. It's just not valuable in terms of advancing the field of AI. [...]That is someone might said it is easy to just install a robot arm, and a pair of camera as eyes for an AI system to achieve end-to-end learning (that is simulate a human constrain), which is completely unrealistic at the current moment.
Oh, I didn't want to suggest that the bot would have to learn how to use a robotic arm. Rather just substituting a signal "move to x" to a SC API with a signal "press right-mouse button at x-coordinate" to a robotic arm.
The use of full map view is also an engineering challenge related to the I/O limitation. Most people don't realize how amazing their "imagination" or "reconstruction" of the field of view really is. For an raw image input to jump all over the map and able to picture them together as a whole and find temporal correlations, and they changed and being covered by units/buildings all the time, is very very difficult.
Remember the Atari engine was trained on pure pixel input. The engine was completely ignorant to the fact that some pixel formed an abstract image of a brick while other pixels formed an abstract image of a ball. I agree that finding temporal correlations is very hard when jumping across screens. But to a bot it's just a bunch of pixels. It doesn't interpret them in any way. Also, even Artosis was able to identify an oracle on the minimap during the cast. The minimap is fully visible at all times and you "just" need to jump in the main screen to zoom in on the action (e.g. placing buildings or microing units).
|
That was amazing. Can't wait to integrate the Deepmind probe saturation into my pvp
|
I wonder where my posts went to. I didn't find them in this thread, how are they merged?
Anyways, if deepmind won't fix the crazy apm, any following games against humans would be meaningless.
|
On January 25 2019 12:45 imp42 wrote:Show nested quote +On January 25 2019 12:14 vesicular wrote:On January 25 2019 09:40 TheDougler wrote:
You don't know that it was the the camera change that actually was the determining factor here. It could be that Mana had a better idea of what he was up against. It could be that the warp prism threw off the AI's gameplan (I think it's this one). It could be that this AI isn't quite as good as other AIs. [...] The final I would say is to play only one agent. Every game used a different agent. It's akin to playing different players. TLO didn't know this when he was playing and played his matches as if it was the same agent and thus tried strats to counter what he just saw in the previous game, which of course didn't work. Playing against a single agent would be quite interesting. A misconception IMO. There is no conceptual difference between "one agent" and "multiple agents", because you can simply combine x agents into one composite agent (which is exactly what they did). Compare it to Innovation switching up his macro game with a 3-rax proxy cheese. It's not akin to playing different players, but the same player choosing a different game plan before he starts the game. The concept of a composite agent gets interesting when you add a super-agent to it, responsible for picking a sub-agent to play a specific game in a boX match. I would imagine the super-agent would then be trained similar to a Texas Hold'em agent and converge to game-theoretical optima for cheese / standard ratio etc.
That's interesting, thanks for sharing. I'm talking more like what you said at the end. Humans know based on how a Bo5 is going when a good time to cheese might be. Or they have a feeling on what build may work based on what they've seen in prior games. This is what I'm talking about. An agent who has a game plan but also adapts their builds based on current events (even if this is only decided prior to the game starting).
|
On January 25 2019 13:06 imp42 wrote:
Oh, I didn't want to suggest that the bot would have to learn how to use a robotic arm. Rather just substituting a signal "move to x" to a SC API with a signal "press right-mouse button at x-coordinate" to a robotic arm.
Still not that easy in practice. First of all, not just the precision of the "move to x" is actually an eye-hand correlation learning task that is the system has to learn how fast the real mouse move relative to the field of view, and the right amount of force to apply to the mouse to stop at the right location related to the screen in physical system. There is a reason why fully automatic robotic factories don't have robotic arms working along side humans, not just for efficiency, but also the precision control of robotics usually meant the force control are fined tuned to a very specific conditions/environments where a little misalignment will crash the whole procedures. From engineering perspective taking all that effort to create such robotics is just not yet cost-effective at the current time. (This kind of precision physical system might take years to reach the same level of proficient as to API command and most likely millions in expenditures just for the equipment).
On January 25 2019 13:06 imp42 wrote:Show nested quote +The use of full map view is also an engineering challenge related to the I/O limitation. Most people don't realize how amazing their "imagination" or "reconstruction" of the field of view really is. For an raw image input to jump all over the map and able to picture them together as a whole and find temporal correlations, and they changed and being covered by units/buildings all the time, is very very difficult. Remember the Atari engine was trained on pure pixel input. The engine was completely ignorant to the fact that some pixel formed an abstract image of a brick while other pixels formed an abstract image of a ball. I agree that finding temporal correlations is very hard when jumping across screens. But to a bot it's just a bunch of pixels. It doesn't interpret them in any way. Also, even Artosis was able to identify an oracle on the minimap during the cast. The minimap is fully visible at all times and you "just" need to jump in the main screen to zoom in on the action (e.g. placing buildings or microing units).
There is a reason why the input of the Atari game is used, due to the limited resolutions of the game and relatively easy to identify patterns in such low resolutions. And actually a lot of types of Atari games are still very poorly performed by machine learning. In order to reconstruct and abstract pixels into objects are not that easy even with CNN networks. The delay and inaccurate results introduced by them will accumulate over time, and difficult to learn. The current raw view is actually just an enhanced minimap hence it is able to tell a lot of information without the actual raw screen full resolution. And the live competition version is still not using the pure raw image input, but as they describe in the blog post
"its perception is restricted to on-screen information, and action locations are restricted to its viewable region.", that is it's attention is limited in the input within an activation zone, and only able to submit commands within that active zone. As I said, if the raw image object recognition is so easy, we will already have all games or problems related to real image solved by machine learning already.
|
On January 25 2019 06:44 Waxangel wrote:Show nested quote +On January 25 2019 06:43 Poopi wrote:
I think the title is kinda misleading.
Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
Let me see the entire map in full detail, let my mind control everything instead of my hands via a clunky keyboard and mouse while I use my eyes to look at a screen, and you say those advantages would be insignificant? You know I have eyelids right? I have to blink, I miss things because of that. I have to scroll the screen around and motion blur messes stuff up. I miss things because of that too.
A human mind would destroy a computer if it had the same inputs and view of the game. It wouldn't be close.
Build a machine that has hands and eyes, make it perform using a monitor, speakers, keyboard and mouse. A human would crush it. Allow the mind to control the game without those clunky devices, and it destroys the AI even more. The mind would be way too fast, way too fast. I can split an army of Blink Stalkers in milliseconds in my mind. It'd be over before it began really.
The human mind is capable of things AI can only dream of. In fact, AI is just one of many achievements of the human mind.Starcraft is a game without perfect information, unlike Chess and Go. It will be a long time before AI comes anywhere close given the same inputs.
|
A few thoughts on various subjects:
APM/EPM: Assuming the APM for AlphaStar is the same as EPM(280), what about for TLO, MaNa, and other pros? I read somewhere it's something around 150 for masters/GM so maybe around 200 for top level pros playing their main race?
Builds/Strats: The commentators were baffled by the heavy emphasis on stalkers, phoenixes, and disruptors. I think it makes perfect sense for an AI that is mechanically better than it's opponents to rely on these units that heavily rewards micro. This would justify the late gases to get more minerals for stalkers. If the AI is pitted against another clone AI for a PvP for 100 games, it would probably adapt by getting a sentry for more efficiency in stalker wars, therefore faster (or rather, more standard) gas timing.
Starcraft2 has historically been balanced for the top 500 or so players; units like high templars are extremely weak in bronze league as players cannot utilize them due to the lack of mechanical skills. So it's obvious that these units with great micro potentials are "overpowered" as they are consistently played inefficiently, even at top human levels. In the current state of the game, I would guess units like immortals, carriers, and tempests (high cost, low micro potential/low speed, low AOE potential, no effect) would not see play in AIvAI gameplay.
Supersaturation: It might be mathematically advantageous to over saturate before expanding, as oppose to getting 16 workers, cut probes while expand earlier, then continue production. Alternatively, could it simply be a buffer against harass tactics human players emphasize that acts as incentive? Again, would it occur in an AIvAI gameplay?
Zerg?: It would be hilarious if AlphaStar 12 pools every game (against humans) simply because of the mechanical advantage.
Edit:
I'm confused about the "difference human and computer input" argument. Isn't having limitations what makes us human? Fine, give them 2 more years to program an AI that has to input commands using mechanical arms, that has to receive information from two local cameras (which shuts off for 400ms every 20s), and don't forget to add blur to the cameras during scrolling: I have a feeling that AlphaStar will still crush it's human counterparts but granted a little bit less of the crushiness. But would the argument turn to the fact that the mechanical arm is made with precise machinery and not tissues? Inversely, if we can play starcraft with just our minds, I'm certain human will still lose. Think: can you go through even 5 simulation in a second?
|
So this bot could micro perfectly given the APM/EPM restrictions on the bot? That's not really fair though because even humans at a specific APM/EPM will make micro mistakes.
|
On January 25 2019 14:29 BronzeKnee wrote:Show nested quote +On January 25 2019 06:44 Waxangel wrote:On January 25 2019 06:43 Poopi wrote:
I think the title is kinda misleading.
Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride. Let me see the entire map in full detail, let my mind control everything instead of my hands via a clunky keyboard and mouse while I use my eyes to look at a screen, and you say those advantages would be insignificant? You know I have eyelids right? I have to blink, I miss things because of that. I have to scroll the screen around and motion blur messes stuff up. I miss things because of that too. A human mind would destroy a computer if it had the same inputs and view of the game. It wouldn't be close. Build a machine that has hands and eyes, make it perform using a monitor, speakers, keyboard and mouse. A human would crush it. Allow the mind to control the game without those clunky devices, and it destroys the AI even more. The mind would be way too fast, way too fast. I can split an army of Blink Stalkers in milliseconds in my mind. It'd be over before it began really. The human mind is capable of things AI can only dream of. In fact, AI is just one of many achievements of the human mind.Starcraft is a game without perfect information, unlike Chess and Go. It will be a long time before AI comes anywhere close given the same inputs.
Point to the cat, please:
![[image loading]](https://savan77.github.io/blog/images/im2.png)
It's almost as though humans and computers are very, very different.
|
I just watched the final game. What a joke the AI was. Massing Blink Stalkers. Moving the entire army back to defend Warp Prism harass. Warping in Stalkers to stop an Immortal drop. Building a cannon to defend against a Warp Prism drop. Not leaving Stalkers in it's base to defend a Warp Prism drop. Blindly charging up ramps with Stalkers versus Immortal/Sentry.
A complete inability to decide how to defend Mana's attack. Without being able to see the entire map, it was terrible.
|
|
|
|
|
|
EPT
