Current Events > Reinforcement Learning already debunks True AI naysayers

Topic List	Page List: 1
Milkman5 05/29/18 4:22:55 AM #1:	with Reinforcement learning, take for instance, an AI wants to play a video game. In the case of Q-learning (a type of reinforcement learning, which is arguably not the most effective reinforcement learning of today, there is feature-based Q-learning etc) AI using Q-learning does not need instructions written by a programmer to make actions it separates every possible unique scenario in the game as a separate state. Every state of the game is assigned a value. The AI can start off with absolutely no value, make random decisions and update the values on what sequence of state led to the greatest reward. Through learning on it's own, updating the possible sequence of states, the AI will eventually converge to the most optimal way to play the game, without any influence of the programmer. There's no reason to believe that this is not exactly how a human thinks fundamentally. Human beings are constantly learning things and changing their actions and behaviors in situations on either what has given them better rewards in the past, or based on arguments others have fed to led them to believe that is the correct action to take. Sure, there are probably some innate human actions that they will default to, but that is also an element of Q-learning. I honestly believe that there is nothing special about human beings and that it's possible to deconstruct our brains as a system of instructions and just because we haven't done so yet, doesn't make it impossible just what people were saying was impossible 20 years ago, ie reinforcement learning, is possible now. Stop thinking that things are impossible just because we haven't figured it out yet. By the same token, obviously I'm making the jump that it will be possible to emulate humans perfectly, but imo Q-learning already does based on what I've observed of humans. I think our 'reward formulas' of each person are just more complicated than those of any currently written rl trained AI <cite>Milkman5 posted [p:902165712] in Reinforcem... [t:76653626]...</cite> <quote>with Reinforcement learning, take for instance, an AI wants to play a video game. In the case of Q-learning (a type of reinforcement learning, which is arguably not the most effective reinforcement learning of today, there is feature-based Q-learning etc) AI using Q-learning does not need instructions written by a programmer to make actions it separates every possible unique scenario in the game as a separate state. Every state of the game is assigned a value. The AI can start off with absolutely no value, make random decisions and update the values on what sequence of state led to the greatest reward. Through learning on it's own, updating the possible sequence of states, the AI will eventually converge to the most optimal way to play the game, without any influence of the programmer. There's no reason to believe that this is not exactly how a human thinks fundamentally. Human beings are constantly learning things and changing their actions and behaviors in situations on either what has given them better rewards in the past, or based on arguments others have fed to led them to believe that is the correct action to take. Sure, there are probably some innate human actions that they will default to, but that is also an element of Q-learning. I honestly believe that there is nothing special about human beings and that it's possible to deconstruct our brains as a system of instructions and just because we haven't done so yet, doesn't make it impossible just what people were saying was impossible 20 years ago, ie reinforcement learning, is possible now. Stop thinking that things are impossible just because we haven't figured it out yet. By the same token, obviously I'm making the jump that it will be possible to emulate humans perfectly, but imo Q-learning already does based on what I've observed of humans. I think our 'reward formulas' of each person are just more complicated than those of any currently written rl trained AI</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:24:48 AM #2:	Yeah because it's working in a domain with clearly defined rules, i.e. a game This does not translate to the real world at all --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902165754] in Reinforcem... [t:76653626]...</cite> <quote>Yeah because it's working in a domain with clearly defined rules, i.e. a game This does not translate to the real world at all</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:27:48 AM #3:	It's just so obvious. Like anyone can see this if they really thought about it. A system of rewards that humans have for example is our pain system. Pain is a negative reward (our nervous system). We do an action we have never done before and feel pain, we learn that it hurts, so now we know this is a less than desirable action. Then we have brain chemicals like dopamine, where we do things and get rewarded by a brain for doing such things. Things that give us this rush, we do again. And we base everything we do on what we think will get us the best outcome, either through trial and error or instructions given by other people. Like 100 people tell you it hurts to break your leg, so now you are convinced of the pain reward so now you know it is not a desirable action to intentionally break your leg. <cite>Milkman5 posted [p:902165804] in Reinforcem... [t:76653626]...</cite> <quote>It's just so obvious. Like anyone can see this if they really thought about it. A system of rewards that humans have for example is our pain system. Pain is a negative reward (our nervous system). We do an action we have never done before and feel pain, we learn that it hurts, so now we know this is a less than desirable action. Then we have brain chemicals like dopamine, where we do things and get rewarded by a brain for doing such things. Things that give us this rush, we do again. And we base everything we do on what we think will get us the best outcome, either through trial and error or instructions given by other people. Like 100 people tell you it hurts to break your leg, so now you are convinced of the pain reward so now you know it is not a desirable action to intentionally break your leg.</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:31:27 AM #4:	What was impossible 20 years ago was finding a way to process a shitton of game states without taking forever to compute or frying your hardware in the process. And that was addressed by clever methods to address a small subset of these states. All of this is still working in a domain where you can iterate through these states and recursively update the values of prior states, and algorithms are able to do this because of the clearly defined rules of whatever game it's playing (and that's how it's able to "play out" and learn from these future states before they actually occur) Everything you just said doesn't have jack shit to do with reinforcement learning --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902165879] in Reinforcem... [t:76653626]...</cite> <quote>What was impossible 20 years ago was finding a way to process a shitton of game states without taking forever to compute or frying your hardware in the process. And that was addressed by clever methods to address a small subset of these states. All of this is still working in a domain where you can iterate through these states and recursively update the values of prior states, and algorithms are able to do this because of the clearly defined rules of whatever game it's playing (and that's how it's able to "play out" and learn from these future states before they actually occur) Everything you just said doesn't have jack shit to do with reinforcement learning</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:35:27 AM #5:	MacadamianNut3 posted... Yeah because it's working in a domain with clearly defined rules, i.e. a game This does not translate to the real world at all I totally disagree. No one has programmed an AI to reinforcement learn Overwatch because there are too many scenarios in the game, so it would take an eternity to learn because with reinforcement learning, the AI would have to play the game for too long to learn to play. But it obviously is possible to train an AI to play Over-watch, just because we don't have the resources currently, doesn't mean it is not theoretically possible. Even with the current Q-learning, the only obstacle is time. The problem is humans were given values at birth, similar to how an AI can be given values at it's birth It would take any person to set the values an eternity, but the values could obviously be set, in this case they were set by evolution or a span of many years. The other issue is, with human learning, we "learn" without playing the game to the end. Because human beings don't have time to Q learn like a computer does. So they learn from the influence of others playing the game currently and in the past and what they observe of others playing the game. So for a computer to replicate humans in that sense, either they would have to replicate them to learn off others in the same way (which seems theoretically doable) or increase the speed of which they can handle different scenarios. <cite>Milkman5 posted [p:902165939] in Reinforcem... [t:76653626]...</cite> <quote><cite>MacadamianNut3 posted...</cite> <quote>Yeah because it's working in a domain with clearly defined rules, i.e. a game This does not translate to the real world at all</quote> I totally disagree. No one has programmed an AI to reinforcement learn Overwatch because there are too many scenarios in the game, so it would take an eternity to learn because with reinforcement learning, the AI would have to play the game for too long to learn to play. But it obviously is possible to train an AI to play Over-watch, just because we don't have the resources currently, doesn't mean it is not theoretically possible. Even with the current Q-learning, the only obstacle is time. The problem is humans were given values at birth, similar to how an AI can be given values at it's birth It would take any person to set the values an eternity, but the values could obviously be set, in this case they were set by evolution or a span of many years. The other issue is, with human learning, we "learn" without playing the game to the end. Because human beings don't have time to Q learn like a computer does. So they learn from the influence of others playing the game currently and in the past and what they observe of others playing the game. So for a computer to replicate humans in that sense, either they would have to replicate them to learn off others in the same way (which seems theoretically doable) or increase the speed of which they can handle different scenarios.</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:43:02 AM #6:	Milkman5 posted... I totally disagree. There is nothing to disagree with. That is exactly how researchers have been able to make advances in that area. It's not a matter of opinion. And it's exactly why the one of the next challenges they are trying to address is superhuman playing ability in Starcraft. Which is different from games such as Chess or Go because there is a lot of process in a real-time and not turn-based system. And even then, this and Overwatch are games where everything you can do is clearly defined in advance. Which once again, doesn't translate to the real world outside of the law of physics. I ignored the rest of the futurology mumbo jumbo in that post. --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166070] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I totally disagree.</quote> There is nothing to disagree with. That is exactly how researchers have been able to make advances in that area. It's not a matter of opinion. And it's exactly why the one of the next challenges they are trying to address is superhuman playing ability in Starcraft. Which is different from games such as Chess or Go because there is a lot of process in a real-time and not turn-based system. And even then, this and Overwatch are games where everything you can do is clearly defined in advance. Which once again, doesn't translate to the real world outside of the law of physics. I ignored the rest of the futurology mumbo jumbo in that post.</quote> ... Copied to Clipboard!
scar the 1 05/29/18 4:45:05 AM #7:	Did you just learn about Q-learning or something? And what's "True AI"? --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166115] in Reinforcem... [t:76653626]...</cite> <quote>Did you just learn about Q-learning or something? And what's "True AI"?</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:45:39 AM #8:	I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. If we are someone still alive when that happens, you have to close your account. <cite>Milkman5 posted [p:902166125] in Reinforcem... [t:76653626]...</cite> <quote>I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. If we are someone still alive when that happens, you have to close your account.</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:46:48 AM #9:	scar the 1 posted... Did you just learn about Q-learning or something? And what's "True AI"? uh no, I didn't. this is based on another topic. By that other topic, "True AI" is AI that does an action without an explicit instruction by a programmer, which is already possible. <cite>Milkman5 posted [p:902166146] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Did you just learn about Q-learning or something? And what's "True AI"?</quote> uh no, I didn't. this is based on another topic. By that other topic, "True AI" is AI that does an action without an explicit instruction by a programmer, which is already possible.</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:49:07 AM #10:	scar the 1 posted... Did you just learn about Q-learning or something? I'm wondering this as well. Q-Learning isn't even new unless you consider something from the 90s to be new so I'm not following the infatuation here --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166181] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Did you just learn about Q-learning or something?</quote> I'm wondering this as well. Q-Learning isn't even new unless you consider something from the 90s to be new so I'm not following the infatuation here</quote> ... Copied to Clipboard!
scar the 1 05/29/18 4:51:16 AM #11:	Milkman5 posted... scar the 1 posted... Did you just learn about Q-learning or something? And what's "True AI"? uh no, I didn't. this is based on another topic. By that other topic, "True AI" is AI that does an action without an explicit instruction by a programmer, which is already possible. Can you link the other topic? (I'm curious) And yeah, if that's what you mean by true AI then we definitely already have it. As for modeling human behavior, we're still a long way off. There are several algorithms that are designed with natural phenomena in mind, like Q-learning, artificial neural networks, genetic algorithms, bird-swarm optimization, etc. It's important to remember, though, that they're models. And if you think a model is ever going to perfectly model reality, then I would disagree with you and think that you're quite naive. --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166221] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Did you just learn about Q-learning or something? And what's "True AI"?</quote> uh no, I didn't. this is based on another topic. By that other topic, "True AI" is AI that does an action without an explicit instruction by a programmer, which is already possible.</quote> Can you link the other topic? (I'm curious) And yeah, if that's what you mean by true AI then we definitely already have it. As for modeling human behavior, we're still a long way off. There are several algorithms that are designed with natural phenomena in mind, like Q-learning, artificial neural networks, genetic algorithms, bird-swarm optimization, etc. It's important to remember, though, that they're models. And if you think a model is ever going to perfectly model reality, then I would disagree with you and think that you're quite naive.</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:52:01 AM #12:	Milkman5 posted... I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. When the fuck did I say it wouldn't be able to I agree that AI will eventually learn to play GAMES. GAMES THAT BY DEFINITION HAVE CLEARLY DEFINED RULES. GAMES =/= THE REAL WURL THAT NO HAF CLEARLY DEFINED RULES SO MAGIC SAUCE Q-LEARNING ISN'T GONNA RESULT IN DOLORES FROM WESTWORLD Bolded to reduce confusion hopefully --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166242] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years.</quote> When the fuck did I say it wouldn't be able to I agree that AI will eventually learn to play GAMES. GAMES THAT BY DEFINITION HAVE CLEARLY DEFINED RULES. GAMES =/= THE REAL WURL THAT NO HAF CLEARLY DEFINED RULES SO MAGIC SAUCE Q-LEARNING ISN'T GONNA RESULT IN DOLORES FROM WESTWORLD Bolded to reduce confusion hopefully</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:53:05 AM #13:	scar the 1 posted... Can you link the other topic? (I'm curious) And yeah, if that's what you mean by true AI then we definitely already have it. As for modeling human behavior, we're still a long way off. There are several algorithms that are designed with natural phenomena in mind, like Q-learning, artificial neural networks, genetic algorithms, bird-swarm optimization, etc. It's important to remember, though, that they're models. And if you think a model is ever going to perfectly model reality, then I would disagree with you and think that you're quite naive. https://gamefaqs.gamespot.com/boards/400-current-events/76653558 <cite>Milkman5 posted [p:902166262] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Can you link the other topic? (I'm curious) And yeah, if that's what you mean by true AI then we definitely already have it. As for modeling human behavior, we're still a long way off. There are several algorithms that are designed with natural phenomena in mind, like Q-learning, artificial neural networks, genetic algorithms, bird-swarm optimization, etc. It's important to remember, though, that they're models. And if you think a model is ever going to perfectly model reality, then I would disagree with you and think that you're quite naive.</quote> https://gamefaqs.gamespot.com/boards/400-current-events/76653558</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:53:55 AM #14:	MacadamianNut3 posted... scar the 1 posted... Did you just learn about Q-learning or something? I'm wondering this as well. Q-Learning isn't even new unless you consider something from the 90s to be new so I'm not following the infatuation here why are you intentionally acting like a jackass? I know you know the context of this topic because you literally posted in the other one <cite>Milkman5 posted [p:902166276] in Reinforcem... [t:76653626]...</cite> <quote><cite>MacadamianNut3 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Did you just learn about Q-learning or something?</quote> I'm wondering this as well. Q-Learning isn't even new unless you consider something from the 90s to be new so I'm not following the infatuation here</quote> why are you intentionally acting like a jackass? I know you know the context of this topic because you literally posted in the other one</quote> ... Copied to Clipboard!
scar the 1 05/29/18 4:54:37 AM #15:	Milkman5 posted... I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. If we are someone still alive when that happens, you have to close your account. Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't. --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166288] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. If we are someone still alive when that happens, you have to close your account.</quote> Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't.</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:56:14 AM #16:	MacadamianNut3 posted... Milkman5 posted... I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years. When the fuck did I say it wouldn't be able to I agree that AI will eventually learn to play GAMES. GAMES THAT BY DEFINITION HAVE CLEARLY DEFINED RULES. GAMES =/= THE REAL WURL THAT NO HAF CLEARLY DEFINED RULES SO MAGIC SAUCE Q-LEARNING ISN'T GONNA RESULT IN DOLORES FROM WESTWORLD Bolded to reduce confusion hopefully The point is how a computer makes it's actions, I'm highlighted the reward based system of making decisions, something that is clearly observable in humans. It's like you are closing your ears and screaming "lalalala" at what I'm saying because scifi writers have made a boogeyman out of AI <cite>Milkman5 posted [p:902166328] in Reinforcem... [t:76653626]...</cite> <quote><cite>MacadamianNut3 posted...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I'll make an account bet with you. That an AI will be able to play professional starcraft play in the next 100 years.</quote> When the fuck did I say it wouldn't be able to I agree that AI will eventually learn to play GAMES. GAMES THAT BY DEFINITION HAVE CLEARLY DEFINED RULES. GAMES =/= THE REAL WURL THAT NO HAF CLEARLY DEFINED RULES SO MAGIC SAUCE Q-LEARNING ISN'T GONNA RESULT IN DOLORES FROM WESTWORLD Bolded to reduce confusion hopefully</quote> The point is how a computer makes it's actions, I'm highlighted the reward based system of making decisions, something that is clearly observable in humans. It's like you are closing your ears and screaming "lalalala" at what I'm saying because scifi writers have made a boogeyman out of AI</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:56:36 AM #17:	Milkman5 posted... I know you know the context of this topic because you literally posted in the other one Context of this topic is given in the opening post. If you're doing some extremely roundabout argument to counter Rika's point that computers can't do anything they are explicitly told to do so, then 1) You're making a shitty argument sry 2) Why the fuck did you feel the need to make a new topic instead of vomiting your drivel in the other one --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166335] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I know you know the context of this topic because you literally posted in the other one</quote> Context of this topic is given in the opening post. If you're doing some extremely roundabout argument to counter Rika's point that computers can't do anything they are explicitly told to do so, then 1) You're making a shitty argument sry 2) Why the fuck did you feel the need to make a new topic instead of vomiting your drivel in the other one</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:57:50 AM #18:	scar the 1 posted... Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't. not yet apparently. I think they released an AI last year and humans were able to beat it <cite>Milkman5 posted [p:902166358] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't.</quote> not yet apparently. I think they released an AI last year and humans were able to beat it</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 4:58:04 AM #19:	Milkman5 posted... The point is how a computer makes it's actions, I'm highlighted the reward based system of making decisions, something that is clearly observable in humans. And you've failed to make the connection to strong AI when all of your examples are about games. Strong AI isn't limited to games, smart one. No shit various AI approaches are attempts at modeling human reasoning. I'm surprised you're not drooling over artificial neural nets --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166361] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>The point is how a computer makes it's actions, I'm highlighted the reward based system of making decisions, something that is clearly observable in humans. </quote> And you've failed to make the connection to strong AI when all of your examples are about games. Strong AI isn't limited to games, smart one. No shit various AI approaches are attempts at modeling human reasoning. I'm surprised you're not drooling over artificial neural nets</quote> ... Copied to Clipboard!
scar the 1 05/29/18 4:58:14 AM #20:	I think you're underestimating the complexity of human consciousness. And I don't wanna put words in Rika's mouth, but it seems like his idea of what "true AI" means is a little more than just "act without explicit human input". --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166364] in Reinforcem... [t:76653626]...</cite> <quote>I think you're underestimating the complexity of human consciousness. And I don't wanna put words in Rika's mouth, but it seems like his idea of what "true AI" means is a little more than just "act without explicit human input".</quote> ... Copied to Clipboard!
Milkman5 05/29/18 4:58:48 AM #21:	MacadamianNut3 posted... Milkman5 posted... I know you know the context of this topic because you literally posted in the other one Context of this topic is given in the opening post. If you're doing some extremely roundabout argument to counter Rika's point that computers can't do anything they are explicitly told to do so, then 1) You're making a shitty argument sry 2) Why the fuck did you feel the need to make a new topic instead of vomiting your drivel in the other one Because any wall of text posted after the first 10 posts in a topic will be ignored by the OP and people who skim the topic. Many will leave the topic thinking that AI needs to be a hard solver to do anything. <cite>Milkman5 posted [p:902166374] in Reinforcem... [t:76653626]...</cite> <quote><cite>MacadamianNut3 posted...</cite> <quote><cite>Milkman5 posted...</cite> <quote>I know you know the context of this topic because you literally posted in the other one</quote> Context of this topic is given in the opening post. If you're doing some extremely roundabout argument to counter Rika's point that computers can't do anything they are explicitly told to do so, then 1) You're making a shitty argument sry 2) Why the fuck did you feel the need to make a new topic instead of vomiting your drivel in the other one</quote> Because any wall of text posted after the first 10 posts in a topic will be ignored by the OP and people who skim the topic. Many will leave the topic thinking that AI needs to be a hard solver to do anything.</quote> ... Copied to Clipboard!
scar the 1 05/29/18 5:00:01 AM #22:	Milkman5 posted... scar the 1 posted... Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't. not yet apparently. I think they released an AI last year and humans were able to beat it Can you provide a link to this, as well? Is it a client-side AI? --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166398] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't.</quote> not yet apparently. I think they released an AI last year and humans were able to beat it</quote> Can you provide a link to this, as well? Is it a client-side AI?</quote> ... Copied to Clipboard!
Milkman5 05/29/18 5:00:58 AM #23:	scar the 1 posted... I think you're underestimating the complexity of human consciousness. And I don't wanna put words in Rika's mouth, but it seems like his idea of what "true AI" means is a little more than just "act without explicit human input". Milkman5 posted... By the same token, obviously I'm making the jump that it will be possible to emulate humans perfectly, but imo Q-learning already does based on what I've observed of humans. I think our 'reward formulas' of each person are just more complicated than those of any currently written rl trained AI I don't think I'm underestimating it. I made sure to state I'm making a jump and assuming a lot of things. It's just what I have observed a humans personally, it's not scientific, but obviously we don't have enough information at the present time to say that is either possible or impossible to replicate human decision making <cite>Milkman5 posted [p:902166414] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote>I think you're underestimating the complexity of human consciousness. And I don't wanna put words in Rika's mouth, but it seems like his idea of what "true AI" means is a little more than just "act without explicit human input".</quote> <cite>Milkman5 posted...</cite> <quote>By the same token, obviously I'm making the jump that it will be possible to emulate humans perfectly, but imo Q-learning already does based on what I've observed of humans. I think our 'reward formulas' of each person are just more complicated than those of any currently written rl trained AI</quote> I don't think I'm underestimating it. I made sure to state I'm making a jump and assuming a lot of things. It's just what I have observed a humans personally, it's not scientific, but obviously we don't have enough information at the present time to say that is either possible or impossible to replicate human decision making</quote> ... Copied to Clipboard!
Milkman5 05/29/18 5:01:58 AM #24:	scar the 1 posted... Milkman5 posted... scar the 1 posted... Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't. not yet apparently. I think they released an AI last year and humans were able to beat it Can you provide a link to this, as well? Is it a client-side AI? I just googled it after you made your post and I found this article. I didn't read the article completely https://www.technologyreview.com/s/609242/humans-are-still-better-than-ai-at-starcraftfor-now/ <cite>Milkman5 posted [p:902166436] in Reinforcem... [t:76653626]...</cite> <quote><cite>scar the 1 posted...</cite> <quote><cite>Milkman5 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't.</quote> not yet apparently. I think they released an AI last year and humans were able to beat it</quote> Can you provide a link to this, as well? Is it a client-side AI?</quote> I just googled it after you made your post and I found this article. I didn't read the article completely https://www.technologyreview.com/s/609242/humans-are-still-better-than-ai-at-starcraftfor-now/</quote> ... Copied to Clipboard!
Milkman5 05/29/18 5:06:02 AM #25:	MacadamianNut3 posted... And you've failed to make the connection to strong AI when all of your examples are about games. Strong AI isn't limited to games, smart one. No s* various AI approaches are attempts at modeling human reasoning. I'm surprised you're not drooling over artificial neural nets if I had a proof then I woudn't be posting it on gamefaqs first. the point is there is no evidence to think it is impossible, and I personally have reasons to lean towards that it possible just in the way human behavior with pain, pleasure, brain chemicals, learning from others and so on. So many people are so clearly living their lives even searching for ways to get high, the next high, the biggest high etc and everything action they take is by design to give them those highs. Like I said, if I had a proof of any of this, then I would be a millionaire and a revolutionary <cite>Milkman5 posted [p:902166530] in Reinforcem... [t:76653626]...</cite> <quote><cite>MacadamianNut3 posted...</cite> <quote>And you've failed to make the connection to strong AI when all of your examples are about games. Strong AI isn't limited to games, smart one. No s* various AI approaches are attempts at modeling human reasoning. I'm surprised you're not drooling over artificial neural nets</quote> if I had a proof then I woudn't be posting it on gamefaqs first. the point is there is no evidence to think it is impossible, and I personally have reasons to lean towards that it possible just in the way human behavior with pain, pleasure, brain chemicals, learning from others and so on. So many people are so clearly living their lives even searching for ways to get high, the next high, the biggest high etc and everything action they take is by design to give them those highs. Like I said, if I had a proof of any of this, then I would be a millionaire and a revolutionary</quote> ... Copied to Clipboard!
scar the 1 05/29/18 5:11:14 AM #26:	Milkman5 posted... scar the 1 posted... Milkman5 posted... scar the 1 posted... Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't. not yet apparently. I think they released an AI last year and humans were able to beat it Can you provide a link to this, as well? Is it a client-side AI? I just googled it after you made your post and I found this article. I didn't read the article completely https://www.technologyreview.com/s/609242/humans-are-still-better-than-ai-at-starcraftfor-now/ Thanks. I don't get the point of your bet, since I think it's the most obvious conclusion in the history of the observable universe that we would be able to make an AI that can beat a human player at Starcraft :). Milkman5 posted... I don't think I'm underestimating it. I made sure to state I'm making a jump and assuming a lot of things. It's just what I have observed a humans personally, it's not scientific, but obviously we don't have enough information at the present time to say that is either possible or impossible to replicate human decision making Don't get me wrong, I think we're making some fascinating progress in terms of replicating human intelligence, but you'll never convince me that we'll be able to perfectly model human behavior. Because models are never perfect. If they were, would they really be models? Just because we know there's an element of a reward mechanism in our brain and we can model a reward mechanism in a computer it doesn't mean that we'll be able to perfectly replicate a human brain. Will we ever make an AI that could pass the Turing test? Debatable. I'm not sure it's even possible by definition to pass that test. When does it end? When can you say with 100% certainty that "you're 100% like a human and nothing you say in the future will make me change my mind"? --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902166633] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote><cite>Milkman5 posted...</cite> <quote><cite>scar the 1 posted...</cite> <quote>Uh, I would think that an AI could already beat a pro in Starcraft. In fact I would be quite surprised if it couldn't.</quote> not yet apparently. I think they released an AI last year and humans were able to beat it</quote> Can you provide a link to this, as well? Is it a client-side AI?</quote> I just googled it after you made your post and I found this article. I didn't read the article completely https://www.technologyreview.com/s/609242/humans-are-still-better-than-ai-at-starcraftfor-now/</quote> Thanks. I don't get the point of your bet, since I think it's the most obvious conclusion in the history of the observable universe that we would be able to make an AI that can beat a human player at Starcraft :). <cite>Milkman5 posted...</cite> <quote>I don't think I'm underestimating it. I made sure to state I'm making a jump and assuming a lot of things. It's just what I have observed a humans personally, it's not scientific, but obviously we don't have enough information at the present time to say that is either possible or impossible to replicate human decision making</quote> Don't get me wrong, I think we're making some fascinating progress in terms of replicating human intelligence, but you'll never convince me that we'll be able to perfectly model human behavior. Because models are never perfect. If they were, would they really be models? Just because we know there's an element of a reward mechanism in our brain and we can model a reward mechanism in a computer it doesn't mean that we'll be able to perfectly replicate a human brain. Will we ever make an AI that could pass the Turing test? Debatable. I'm not sure it's even possible by definition to pass that test. When does it end? When can you say with 100% certainty that "you're 100% like a human and nothing you say in the future will make me change my mind"?</quote> ... Copied to Clipboard!
Milkman5 05/29/18 5:17:42 AM #27:	if we can't then I think it will be because of a limitation we have as thinkers and not that the model can't be made. (or it is because we died off before we could) But I guess you're right in the sense I have no reason to believe it's obviously the case that humans function the same way. I'm just letting my gut feelings get the best of me. The point should have been that there's no convincing argument or no proof that humans behave in some unique way that could never be replicated, and we already have ways to make a system of instructions that is more than just do this because I said so. <cite>Milkman5 posted [p:902166762] in Reinforcem... [t:76653626]...</cite> <quote>if we can't then I think it will be because of a limitation we have as thinkers and not that the model can't be made. (or it is because we died off before we could) But I guess you're right in the sense I have no reason to believe it's obviously the case that humans function the same way. I'm just letting my gut feelings get the best of me. The point should have been that there's no convincing argument or no proof that humans behave in some unique way that could never be replicated, and we already have ways to make a system of instructions that is more than just do this because I said so.</quote> ... Copied to Clipboard!
MacadamianNut3 05/29/18 5:22:45 AM #28:	Milkman5 posted... if I had a proof then I woudn't be posting it on gamefaqs first. I didn't say proof. I said a connection Your argument is basically boiling down to "Q-Learning mimics how humans behave and interact with the world, therefore someday AI should be possible to mimic humans exactly and be considered as sentient as humans because both humans and AI would be doing the same thing" Only problem there is that as I said earlier, mimicking human behavior is the whole motivation behind artificial intelligence, hence the name. So reinforcement learning working well in games doesn't hint at anything in regards to strong AI just because it mimics human behavior. A lot of AI approaches mimic human behavior. Being able to play chess, go, starcraft or whatever has as much to do with strong AI as a neural network being able to recognize cat images. --- Roll Tide & Go Irish <cite>MacadamianNut3 posted [p:902166848] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>if I had a proof then I woudn't be posting it on gamefaqs first.</quote> I didn't say proof. I said a connection Your argument is basically boiling down to "Q-Learning mimics how humans behave and interact with the world, therefore someday AI should be possible to mimic humans exactly and be considered as sentient as humans because both humans and AI would be doing the same thing" Only problem there is that as I said earlier, mimicking human behavior is the whole motivation behind artificial intelligence, hence the name. So reinforcement learning working well in games doesn't hint at anything in regards to strong AI just because it mimics human behavior. A lot of AI approaches mimic human behavior. Being able to play chess, go, starcraft or whatever has as much to do with strong AI as a neural network being able to recognize cat images.</quote> ... Copied to Clipboard!
scar the 1 05/29/18 5:42:30 AM #29:	Milkman5 posted... if we can't then I think it will be because of a limitation we have as thinkers and not that the model can't be made. (or it is because we died off before we could) But I guess you're right in the sense I have no reason to believe it's obviously the case that humans function the same way. I'm just letting my gut feelings get the best of me. The point should have been that there's no convincing argument or no proof that humans behave in some unique way that could never be replicated, and we already have ways to make a system of instructions that is more than just do this because I said so. Well I mean we haven't been able to perfectly replicate anything with a model, because that's not what models do. --- Everything has an end, except for the sausage. It has two. <cite>scar the 1 posted [p:902167198] in Reinforcem... [t:76653626]...</cite> <quote><cite>Milkman5 posted...</cite> <quote>if we can't then I think it will be because of a limitation we have as thinkers and not that the model can't be made. (or it is because we died off before we could) But I guess you're right in the sense I have no reason to believe it's obviously the case that humans function the same way. I'm just letting my gut feelings get the best of me. The point should have been that there's no convincing argument or no proof that humans behave in some unique way that could never be replicated, and we already have ways to make a system of instructions that is more than just do this because I said so.</quote> Well I mean we haven't been able to perfectly replicate anything with a model, because that's not what models do.</quote> ... Copied to Clipboard!
YOUHAVENOHOPE 05/29/18 5:49:05 AM #30:	itt: a very smart person argues about AI with someone with a PHD in a related field --- https://imgur.com/kpXTakw https://imgur.com/Pto3yPa http://media.giphy.com/media/NBVBnHv8bV13q/giphy.gif <cite>YOUHAVENOHOPE posted [p:902167304] in Reinforcem... [t:76653626]...</cite> <quote>itt: a very smart person argues about AI with someone with a PHD in a related field</quote> ... Copied to Clipboard!
COVxy 05/29/18 5:51:47 AM #31:	Reinforcement learning was studied in psychology and neuroscience long before AI. This revelation that this algorithm may be how the brain works is silly, since you are literally working in reverse. --- =E[(x-E[x])(y-E[y])] <cite>COVxy posted [p:902167364] in Reinforcem... [t:76653626]...</cite> <quote>Reinforcement learning was studied in psychology and neuroscience long before AI. This revelation that this algorithm may be how the brain works is silly, since you are literally working in reverse.</quote> ... Copied to Clipboard!
Topic List	Page List: 1

Current Events > Reinforcement Learning already debunks True AI naysayers (+) Yeah! (+)

Current Events > Reinforcement Learning already debunks True AI naysayers