|
noooooo -> RE: Concerning the AI (11/18/2018 8:30:28 PM)
|
quote:
ORIGINAL: vorbarra2
That's true, but not useful. You can teach a computer how to recognize a cat with that type of learning, and therefore any other type of image recognition, but thats a binary outcome. Is the image cat? is a yes/no question. In the case of Go, though the board state is complex, there are still only a finite number of states the game qcan can exist in, snd the only imput space for the AI is 'where do I put the counter?'
AI for games with an essentially unbounded number of internal states (ie. all computer game design as we know it) does not conform to this type of learning. The number of instructions the AI has to process is very large, it's not 'the counter goes here,' but many different types of instructions which don't map neatly to the binary outcome classification of machine learning. The best that can be achieved is a generalized algorithim, which probablys outperforms the majority of human players (certainly if my record is anythign to go by) The comments by the dev are trying to convey this point, which is basically 'so how would you express that in a symbolic, generalized fashion that would apply to all possible edge states and conditions and unit choices?'
Things like pathfinding and pathing though are more solvable and I think the way in which units path could certainly be improved. It;s not that obvious to me why they take the path they do, and the interface does a poor job of telegraphing what commands would be needed to ensure a certain type of movement. I can formulate the thoguht 'mech inf, stick to roads, get to the town fast, chuck out your passnegers and take up defesnive positions in or behind buildings,' but the pathing or formation commands never seem to make it easy to achieve this.
I wonder if the comments people are making about piecemeal advances are really an issue of tactical-level decision making, or simply poor pathing leading to formations getting broken up and encountered piecemeal? Does the AI have an internal metric for checking formation integrity?
Took the words right out of my mouth. When you look a go board, say a 19x19 board, you have 361 possible places to put your stones. And less over time as the board becomes filled. Even that creates a enormous amount of possible permutations which is why AlphaZero (chess AI although chess has much less permutations than Go) and AlphaGo is so impressive.
But if you look at something like AB or other computer games, the amount of possible states outnumber Go by such a massive amount that it's not even close to comparable. The fact is, when a tank is pointing 56 degrees northeast is technically a different state than if everything else being the same the tank point 57 degrees. This is talking about ONE tank and ONLY the direction it's facing. It's simply not even remotely on the same magnitude as something like Go, and Go is ALREADY a massive undertaking.
Using ML for something like this is not the equivalent of recognizing a cat or playing Go. It's like asking it to speak English to the level of an average human. There are so many contexts and concepts that there simply isn't enough computing power or resources to do it.
|
|
|
|
