|
Post by mystic spastic on Aug 18, 2019 12:59:50 GMT -5
I know what it took for computers to beat humans at Chess, Jeopardy and Go. 1) Start with hardware that takes up a room
Any idea what it would require, compute-wise, to make a realistic NPC opponent/teammate in War Robots? An NPC that would not break immersion, and used equipment equal in speed/HP/damage to human players.
|
|
|
Post by linearblade on Aug 18, 2019 16:26:47 GMT -5
I know what it took for computers to beat humans at Chess, Jeopardy and Go. 1) Start with hardware that takes up a room
Any idea what it would require, compute-wise, to make a realistic NPC opponent/teammate in War Robots? An NPC that would not break immersion, and used equipment equal in speed/HP/damage to human players.
A human like ai? Probably difficult. A computer one? No very hard, they’ve been making ai for years. Think every pc / Xbox shooter fir the last 2 decades.m or more. To make a learning ai, aka neural net. Probably not difficult if they had the technology. In essence, you could train any ai with just the game inputs alone, and run enough trainers till you got a hit. It would start slow because some bots would end up being duds etc. but eventually they would learn shooting reds instead of blues etc was good. Or calling beacons was good. Another way would be to simply train it by watching replays. This is how they did Starcraft. With all that said. It’s likely those ai would be very alien. Because they would behave like a computer does. That is, being able to process many more simultaneous inputs/outputs than a human can. So to make an immersive, human alike ai - difficult A basic ai - trivial(ish) - if they hired the right guy A strong ai - probably still trivial A true neural net / other algorithm designed to learn... probably impossible to very difficult if they lack the skill set. Otherwise moderately difficult
|
|
|
Post by 乃ㄥ卂乙 on Aug 18, 2019 16:51:30 GMT -5
Maybe that's what's going on with those low level account sightings in Champions League. Maybe pixo is using those accounts to train/develop new ai :/
|
|
|
Post by mystic spastic on Aug 18, 2019 20:52:06 GMT -5
Maybe that's what's going on with those low level account sightings in Champions League. Maybe pixo is using those accounts to train/develop new ai :/ Could very well be.
|
|
|
Post by oooohmyyy on Aug 18, 2019 21:19:03 GMT -5
I know what it took for computers to beat humans at Chess, Jeopardy and Go. 1) Start with hardware that takes up a room
Any idea what it would require, compute-wise, to make a realistic NPC opponent/teammate in War Robots? An NPC that would not break immersion, and used equipment equal in speed/HP/damage to human players.
Since this game only uses simple left and right camera view, Id say its pretty easy to visualize each map from birdview and make it a 2 dimensional graph, draw the surrounding area of each strategic point(a beacon, ramp, cover, snipe point etc.) into a node, and have each node labelled in x y coordinates, therefore we can visualize each teammate and enemies location as a point in a node area. We have our parameters of the machine learning database. They can therefore collect veteran players gameplays and convert them into moving dots on 2D node graphs, and train using neural networks to make a War Robots AI that behaves in a way similar to veteran players and able to counter various situations dynamically.
|
|
|
Post by linearblade on Aug 18, 2019 22:21:06 GMT -5
I know what it took for computers to beat humans at Chess, Jeopardy and Go. 1) Start with hardware that takes up a room
Any idea what it would require, compute-wise, to make a realistic NPC opponent/teammate in War Robots? An NPC that would not break immersion, and used equipment equal in speed/HP/damage to human players.
Since this game only uses simple left and right camera view, Id say its pretty easy to visualize each map from birdview and make it a 2 dimensional graph, draw the surrounding area of each strategic point(a beacon, ramp, cover, snipe point etc.) into a node, and have each node labelled in x y coordinates, therefore we can visualize each teammate and enemies location as a point in a node area. We have our parameters of the machine learning database. They can therefore collect veteran players gameplays and convert them into moving dots on 2D node graphs, and train using neural networks to make a War Robots AI that behaves in a way similar to veteran players and able to counter various situations dynamically. If they were gonna train it, it wouldn’t really matter about the 2s / 3d The thing about ai, which in my eyes is sort of hocus pocus Sham, Is that much of the ai itself is blackbox. The ai is a tool that “Learns” but it’s code itself is often out of the range of understanding of the human parent, usually because the steps the ai gets there with may be convoluted or alien is design
|
|
|
Post by mystic spastic on Aug 18, 2019 22:32:30 GMT -5
I know what it took for computers to beat humans at Chess, Jeopardy and Go. 1) Start with hardware that takes up a room
Any idea what it would require, compute-wise, to make a realistic NPC opponent/teammate in War Robots? An NPC that would not break immersion, and used equipment equal in speed/HP/damage to human players.
Since this game only uses simple left and right camera view, Id say its pretty easy to visualize each map from birdview and make it a 2 dimensional graph, draw the surrounding area of each strategic point(a beacon, ramp, cover, snipe point etc.) into a node, and have each node labelled in x y coordinates, therefore we can visualize each teammate and enemies location as a point in a node area. We have our parameters of the machine learning database. They can therefore collect veteran players gameplays and convert them into moving dots on 2D node graphs, and train using neural networks to make a War Robots AI that behaves in a way similar to veteran players and able to counter various situations dynamically. What I'm really wondering about, is the amount of compute power needed to do this real time in a 6v6 PVE game. I'm assuming the compute power would need to be on the back end and not on the mobile device itself. Could mere CPUs do it? TPUs? How many TPUs per AI player?
I should go play some BoT again. See if it's immediately discernible which bots are run by NPC. I know I typically outscored them, but don't recall thinking - oh, that bot's being run by a "bot".
It was just too easy to locate and dominate the NPCs in Mech Battle, just as the NPCs die like flies in PUBG.
|
|
|
Post by mystic spastic on Aug 18, 2019 23:05:45 GMT -5
Since this game only uses simple left and right camera view, Id say its pretty easy to visualize each map from birdview and make it a 2 dimensional graph, draw the surrounding area of each strategic point(a beacon, ramp, cover, snipe point etc.) into a node, and have each node labelled in x y coordinates, therefore we can visualize each teammate and enemies location as a point in a node area. We have our parameters of the machine learning database. They can therefore collect veteran players gameplays and convert them into moving dots on 2D node graphs, and train using neural networks to make a War Robots AI that behaves in a way similar to veteran players and able to counter various situations dynamically. If they were gonna train it, it wouldn’t really matter about the 2s / 3d The thing about ai, which in my eyes is sort of hocus pocus Sham, Is that much of the ai itself is blackbox. The ai is a tool that “Learns” but it’s code itself is often out of the range of understanding of the human parent, usually because the steps the ai gets there with may be convoluted or alien is design IME, while the math behind the code is not trivial, someone with a little calculus and algebra knowledge can get a fair understanding on what's happening with the various activation functions and back propogation algorithms. At least that's my experience so far. But I'm only familiar with fairly simple net algorithms, not having ventured into LTSM or convolutional.
What's alien, are the numbers for bias and weight that the neural network eventually assigns to each node through training. They don't seem to "make sense" to us humans.
Second random (pun unintended) question of the day - given that a nodes initial weights are assigned randomly*, just what effect does that have on final, trained biases and weights for a given neural net and training data? Given the same training set, do all equal NN's optimize identically, regardless of the randomness of their initial node weights?
Answer: Apparently not. The final trained nets do perform similarly, though not identically...
* Yes, I know pseudorandom, and about He and Xavier initialization methods.
|
|
|
Post by linearblade on Aug 19, 2019 1:25:14 GMT -5
If they were gonna train it, it wouldn’t really matter about the 2s / 3d The thing about ai, which in my eyes is sort of hocus pocus Sham, Is that much of the ai itself is blackbox. The ai is a tool that “Learns” but it’s code itself is often out of the range of understanding of the human parent, usually because the steps the ai gets there with may be convoluted or alien is design IME, while the math behind the code is not trivial, someone with a little calculus and algebra knowledge can get a fair understanding on what's happening with the various activation functions and back propogation algorithms. At least that's my experience so far. But I'm only familiar with fairly simple net algorithms, not having ventured into LTSM or convolutional.
What's alien, are the numbers for bias and weight that the neural network eventually assigns to each node through training. They don't seem to "make sense" to us humans.
Second random (pun unintended) question of the day - given that a nodes initial weights are assigned randomly*, just what effect does that have on final, trained biases and weights for a given neural net and training data? Given the same training set, do all equal NN's optimize identically, regardless of the randomness of their initial node weights?
Answer: Apparently not. The final trained nets do perform similarly, though not identically...
* Yes, I know pseudorandom, and about He and Xavier initialization methods.
To focus in on the alien part Yes, we can read the code generated an trace through it. But we won’t understabd the reasoning for it. It’s not going to be an evolutionary impulse that we would see in an animal. It will be evolution because “this was better” since part of many algorithms are brute force , there may be no logic other than, this bit changed and it was better. Therefore it was selected. So understanding why the 2 steps backward gets you 10 steps forward part may not be understood. Thus alien. Additionally, much of what AI do is simply down to how many click it can execute in a given time. If we take away the apm aspect, these games get much closer. In a game like war robots, I have to wonder how efficient it’s able to truly be. Yes it can corner shoot and step perfectly. And this might be the ultimate edge. But oftentimes it’s just efficient trading of units that win games. And corner shooting bots taking perfect smiles don’t win games. They get spawn raided. It’s an interesting subject , and I have to assume any game with little change can be “solved”, but the reality is. An AI is still trained to do something very specific. Abs if you change the inputs it may not be able to compensate Even games like Starcraft. A human player given to thought could probably break the AI because AI are typically reactionary instead of proactive In the end it just goes to show: “you can’t beat the system, unless you change the rules”
|
|