antor
Junior Member
Posts: 87
Likes: 51
Country: Sweden
Politics: Middle Left something
Religion: Apatheist
Age: 35
|
Post by antor on May 15, 2021 13:58:14 GMT
I have for a long time been preparing to code a game or conflict model that captures at least part of human behavior.
In Chess and Go there are now neural network based engines that beat grandmasters easily. And the engines play against eachother too, with ofter interesdting results.
Now my aim is to reduce the actual game in between these engines to a minimum. Simply for puritys sake. I went on a long thought trip on this and thought about Tug of war which is 1 dimensional, 1 variable. Simple right? Nah wanna go simpler, because real tug of war needs two points of contact and contact with the ground and grip matters. Next then, anyone heard of magic the gathering? That's become my new inspiration.
Attack and Defense. Player A and B set Attack and Defense numbers. Then fight. If A guessed the Defense of B to 3, it would of course set its attack to 3. If B defence actually was exactly 3 then A guessed spot on and A gets 100% reward on the attack.
But how to guess? Looking directly at the opponents neural net. Each (cpu) player has neuron inputs directly tied to the opponent neural net. How to best lay this net out geometrically I havent figured out yet.
But basically. The players "see" eachother (eachoters net which is like seeing eachothers code) and who can best approximate eachothers code is who would win, ideally. Its a simple principle that we use all the time. If I "know" you it means I have some kind of approximation of your behaviour in my head. It can be used for evil but also good in the form of cooperation. Now if I can make an approximation of a new person both fast and accurate then I should be very good at this game.
|
|
antor
Junior Member
Posts: 87
Likes: 51
Country: Sweden
Politics: Middle Left something
Religion: Apatheist
Age: 35
|
Post by antor on May 15, 2021 14:04:58 GMT
Also, about the "reward" on the conflict. I was thinking to actually add noise on neural net inputs as a punishment. So theres no actual reward just less punishment in that case. It fits well into the thought model. I mean if i hit extremely hard it is going make me more exhausted and less able to focus in the near term for the rest of the fight. This could be modeled in the game as simple noise on the "eye-neurons" of the player.
|
|
|
|
Post by Eugene 2.0 on May 17, 2021 17:34:32 GMT
Unfortunately, I haven't realized the logic of this game. To forecast the cpu's moves (here 'cpu' is the enemy) is to know its strategy, i.e. it's code, right? So, the game is to deciphe each other's programmed moves sequence before you making another step - as it is in Chess or Go.
Well, I'm not an expert for strategies like those, because I'm definitely not a warrior.
Anyway, I think that a little reward is a good point indeed, while I think that it could be the general lack of pursuing to be a leader. I mean as Plato's wrote in nis very first book of "Republica": nobody wants to be a ruler (i.e. a leader).
Anyone who wants to become a leader probably has something like a talent or some skills that help him to overcome some difficulties. I've lost all the courage to win anything, and that's why there's no point for me to be a winner or a leader. But I don't really know how it goes in machines. How AI responses to it? Do it have something as a wish to conquer the world?
|
|
antor
Junior Member
Posts: 87
Likes: 51
Country: Sweden
Politics: Middle Left something
Religion: Apatheist
Age: 35
|
Post by antor on May 17, 2021 17:57:07 GMT
Unfortunately, I haven't realized the logic of this game. To forecast the cpu's moves (here 'cpu' is the enemy) is to know its strategy, i.e. it's code, right? So, the game is to deciphe each other's programmed moves sequence before you making another step - as it is in Chess or Go. Well, I'm not an expert for strategies like those, because I'm definitely not a warrior. Anyway, I think that a little reward is a good point indeed, while I think that it could be the general lack of pursuing to be a leader. I mean as Plato's wrote in nis very first book of "Republica": nobody wants to be a ruler (i.e. a leader). Anyone who wants to become a leader probably has something like a talent or some skills that help him to overcome some difficulties. I've lost all the courage to win anything, and that's why there's no point for me to be a winner or a leader. But I don't really know how it goes in machines. How AI responses to it? Do it have something as a wish to conquer the world? I think you realized the logic well enough  The neural net configuration is the same as "code" in this case. So the entity which is best at deciphering the opponent behaviour from this code, is best at predicting the opponent and therefore wins. But, thats only one scenario. Some AI may develop a strategy to focus only on a small part of the opponent code, make a crude but fast strategy. Time is a factor too, unless I make the game turn based. Anyway I believe there will be a balance there, where several strategies are possible. But I dont know what will happen exactly, I want to find out though thats why I'm doing this. "How AI responses to it? Do it have something as a wish to conquer the world?" I'm not actually interested in so big questions in this project. But just to mention, one key thing about AI is it doesn't understand. In the same way. It's much more likely that AI would conquer the world based on a misunderstanding or a byproduct of its masters wishes. Google "paperclip ai". But anyways I believe in a hybrid future. Ai has already conquered our pockets so who's to claim it hasn't conquered us all. |
|
|
|
Post by Eugene 2.0 on May 17, 2021 19:26:23 GMT
Unfortunately, I haven't realized the logic of this game. To forecast the cpu's moves (here 'cpu' is the enemy) is to know its strategy, i.e. it's code, right? So, the game is to deciphe each other's programmed moves sequence before you making another step - as it is in Chess or Go. Well, I'm not an expert for strategies like those, because I'm definitely not a warrior. Anyway, I think that a little reward is a good point indeed, while I think that it could be the general lack of pursuing to be a leader. I mean as Plato's wrote in nis very first book of "Republica": nobody wants to be a ruler (i.e. a leader). Anyone who wants to become a leader probably has something like a talent or some skills that help him to overcome some difficulties. I've lost all the courage to win anything, and that's why there's no point for me to be a winner or a leader. But I don't really know how it goes in machines. How AI responses to it? Do it have something as a wish to conquer the world? I think you realized the logic well enough :) The neural net configuration is the same as "code" in this case. So the entity which is best at deciphering the opponent behaviour from this code, is best at predicting the opponent and therefore wins. But, thats only one scenario. Some AI may develop a strategy to focus only on a small part of the opponent code, make a crude but fast strategy. Time is a factor too, unless I make the game turn based. Anyway I believe there will be a balance there, where several strategies are possible. But I dont know what will happen exactly, I want to find out though thats why I'm doing this. "How AI responses to it? Do it have something as a wish to conquer the world?" I'm not actually interested in so big questions in this project. But just to mention, one key thing about AI is it doesn't understand. In the same way. It's much more likely that AI would conquer the world based on a misunderstanding or a byproduct of its masters wishes. Google "paperclip ai". But anyways I believe in a hybrid future. Ai has already conquered our pockets so who's to claim it hasn't conquered us all. AI might concentrate on some pieces of data to approximate it, and then, as I presume, it chooses the tactics to put some obstacles or barriers to finish the strategy, right? I mean this is a process, and all of it's going on a real time, isn't it? I wouldn't say that AI had conquered, but some people gave up to struggle this process, or they just followed that conquerement. It's like people have given up before the temptations of malls, coffee machines, phones, and so on - the self-destruct instinct, I presume. What do you mean by a balance process in that game? To balance what? |
|
antor
Junior Member
Posts: 87
Likes: 51
Country: Sweden
Politics: Middle Left something
Religion: Apatheist
Age: 35
|
Post by antor on May 19, 2021 8:43:25 GMT
I think you realized the logic well enough The neural net configuration is the same as "code" in this case. So the entity which is best at deciphering the opponent behaviour from this code, is best at predicting the opponent and therefore wins. But, thats only one scenario. Some AI may develop a strategy to focus only on a small part of the opponent code, make a crude but fast strategy. Time is a factor too, unless I make the game turn based. Anyway I believe there will be a balance there, where several strategies are possible. But I dont know what will happen exactly, I want to find out though thats why I'm doing this. "How AI responses to it? Do it have something as a wish to conquer the world?" I'm not actually interested in so big questions in this project. But just to mention, one key thing about AI is it doesn't understand. In the same way. It's much more likely that AI would conquer the world based on a misunderstanding or a byproduct of its masters wishes. Google "paperclip ai". But anyways I believe in a hybrid future. Ai has already conquered our pockets so who's to claim it hasn't conquered us all. AI might concentrate on some pieces of data to approximate it, and then, as I presume, it chooses the tactics to put some obstacles or barriers to finish the strategy, right? I mean this is a process, and all of it's going on a real time, isn't it? I wouldn't say that AI had conquered, but some people gave up to struggle this process, or they just followed that conquerement. It's like people have given up before the temptations of malls, coffee machines, phones, and so on - the self-destruct instinct, I presume. What do you mean by a balance process in that game? To balance what? Yeah obstacles (the block factor) is part of the game, like for any game. But theres ofc attack as well, of equal importance. Real time or turn based, I dont know. Magic the gathering is turn based. But I dont know which one to try out first, probably just whatever looks the simplest. Game design isn't easy Balance the players. Cpu1 and cpu2 will play and evolve their strategies. Which means they evolve their neural nets in the process. Im just interested in what that will actually look like, what happens. |
|
|
|
Post by karl on Jun 21, 2021 0:58:19 GMT
I have for a long time been preparing to code a game or conflict model that captures at least part of human behavior. In Chess and Go there are now neural network based engines that beat grandmasters easily. And the engines play against eachother too, with ofter interesdting results. Now my aim is to reduce the actual game in between these engines to a minimum. Simply for puritys sake. I went on a long thought trip on this and thought about Tug of war which is 1 dimensional, 1 variable. Simple right? Nah wanna go simpler, because real tug of war needs two points of contact and contact with the ground and grip matters. Next then, anyone heard of magic the gathering? That's become my new inspiration. Attack and Defense. Player A and B set Attack and Defense numbers. Then fight. If A guessed the Defense of B to 3, it would of course set its attack to 3. If B defence actually was exactly 3 then A guessed spot on and A gets 100% reward on the attack. But how to guess? Looking directly at the opponents neural net. Each (cpu) player has neuron inputs directly tied to the opponent neural net. How to best lay this net out geometrically I havent figured out yet. But basically. The players "see" eachother (eachoters net which is like seeing eachothers code) and who can best approximate eachothers code is who would win, ideally. Its a simple principle that we use all the time. If I "know" you it means I have some kind of approximation of your behaviour in my head. It can be used for evil but also good in the form of cooperation. Now if I can make an approximation of a new person both fast and accurate then I should be very good at this game.
This is a challenge I face in a card game I'm currently coding. If player1 has a particular hand A and player2 has a particular hand B, then player1 can make great gain by making choice X, but if player2 knows the move will be made, it will make a counter move. So how to code the AI of player1? If it always makes choice X, one will eventually learn to always make the counter move. If it never makes it, it will always miss out on the chance of making great gain. So the solution was to code randomness into it. It will often make choice X in that given situation, but not always.
|
|
antor
Junior Member
Posts: 87
Likes: 51
Country: Sweden
Politics: Middle Left something
Religion: Apatheist
Age: 35
|
Post by antor on Jun 21, 2021 14:52:30 GMT
Yes in your case it was a problem. A chicken/egg situation between the players. Here I want to make that the main part of the game. Maybe it will allow too many different strategies for it to be any fun but I'll try anyway.
|
|
The neural net configuration is the same as "code" in this case. So the entity which is best at deciphering the opponent behaviour from this code, is best at predicting the opponent and therefore wins. But, thats only one scenario. Some AI may develop a strategy to focus only on a small part of the opponent code, make a crude but fast strategy. Time is a factor too, unless I make the game turn based. Anyway I believe there will be a balance there, where several strategies are possible. But I dont know what will happen exactly, I want to find out though thats why I'm doing this.
