Could a computer be good at playing ASL?

[bendizoid]

The smarter computers get the dumber we get. This race for GAI will be our downfall.

 

[Sparafucil3]

Codifying Andrea Fantozzi's Play By Self (https://www.facebook.com/groups/358520268920548/permalink/1078447893594445) probably would be the best chance at getting something that would play a reasonable game. It would be a big task, but something like this I think is better suited than the smaller decision space of chess/go (which are of course huge in of themselves) - but Chess/Go's ease of ascertaining the game-score at any given point is many orders of magnitudes easier than ASL's so for me you'd need more of a programmed decision tree rather than a neural-net style approach.

When you break it down, there are so many types of scenarios in war, which are modeled in ASL. Meeting engagement, attack against a hasty defense, attack against a prepared defense, breaching operations, point defense, defense in depth, movement to contact, fighting withdrawal, etc. With proper reinforcement learning, AI could possibly give you a good game but it would take a lot of make sure the reinforcement model was properly modeled. Certainly, if the playing field were constrained to one board (like go) it would be far easier to build the specific model. If you wanted a general model, it would take countless games across countless boards to get something good. -- jim

 

[PresterJohn]

Shall we play a game?

1. Chess
2. Checkers
3. Advanced Squad Leader
4. Global Thermonuclear War

 

[zgrose]

If you wanted a general model, it would take countless games across countless boards to get something good. -- jim

im not sure you understand the basic premise of AI given this statement. That being said, yes, all AI, of any tier, require training. as do humans.

 

[PresterJohn]

This is one of those cases where, if you are looking for a Wargame Opponent Player Replacement, it really will be the DoD that will have the most insight.

 

[Sparafucil3]

im not sure you understand the basic premise of AI given this statement. That being said, yes, all AI, of any tier, require training. as do humans.

I have made, trained, productized, and kept current more AI/ML models than most people around these parts. One of my real-world applications keeps people safe.

When I say "general" I mean non-board and non-scenario specific. Go is a single scenario and a single board. But to train an AI model to generalize a fighting withdrawal on maps it has never seen would be something more. Imagine Alpha-go playing Go on a board it has never seen before. Imagine Alpha-go playing with pieces that vary from one game to the next, where the movement and capture capabilities vary from game to game. That's what I mean when I say general model. Alpha-go is VERY good at a VERY specific game that doesn't change from one playing to the next.

WRT training, humans are better adapted to generalize their learning from one set of conditions and extrapolate to another. Our knowledge has become generally applicable. AI's learning it constrained to its environment. Add another row of squares to the center of a chess board and everything AI knows lacks context. A human would adapt more quickly. In the long run, our robot overlords would again rule the roost. -- jim

 

[von Marwitz]

It would be interesting to set up machine learning for an AI to play ASL, but first the rules would have to be clearly defined.

And a great database would be needed, which is not there.

I can also imagine that it would be difficult to keep an AI from "cheating", meaning that during play, human players regularly make rules mistakes or overlook them. An AI would "learn" that and who could reliably catch it doing so?

von Marwitz

 

[PresterJohn]

In my understanding, if the rules are clearly defined (coded) then the AI can't break them. Once that is done the AI gets to move pieces and roll dice a few million (or maybe billion) times until it gets some idea of how to win, or at least not lose badly.

 

[ScottRomanowski]

I think @von Marwitz has a good point. Machine learning takes examples and creates its own rules. If the examples include sometimes overlooking some rules, the machine will learn that behavior. The plot of Fred Saberhagen's novel "Octagon" is based on something similar.

 

[zgrose]

AGI probably won’t be trained in the same way the current generation of AIs are trained.

 

[Blaze]

Computers are good at playing chess because chess is principally a game of calculation. If a computer was powerful enough to play ASL do you think it could be good at it? Is ASL also a game of calculation or does some sort of artistic or intuitive skill come into it?

Yes! The Technology has been with us since 1984...
Picture evidence below:

Computing 6 double threes. Computing possibilities.
RESPONCE: "Multiple Hits" Burning AFV "Hasta la vista, baby"

 

[Horrido]

The only way it will happen in our lifetime is if regenerative AI becomes less error prone. That is a HUGE hurtle, and I have my doubts it will be solved anytime soon. The old method of decision tree AI would work, but would take too much human manpower to accomplish. If you used the old decision tree method to control the rules, and used regenerative AI to run millions of iterations of each scenario... Now you are talking. But even Kurt couldn't afford that. Maybe if he would have invested that Tabula Rossa money. The catch 22 is that if its smart enough to play a scenario well, it would be smart enough to learn the rules. Unfortunately the intelligence part of regenerative AI is an illusion. It's very useful, but always take something like chatGPT with a grain of salt. That being said, I use it every day. Its just so darn fun to argue with a robot.

 

[PresterJohn]

My version is almost ready. I just need a business partner with a suitable stature in the hobby.

 

[von Marwitz]

But even Kurt couldn't afford that.

We need to get Jeff Bezos and Elon Musk to play ASL...

von Marwitz

 

[Actionjick]

And a great database would be needed, which is not there.

I can also imagine that it would be difficult to keep an AI from "cheating", meaning that during play, human players regularly make rules mistakes or overlook them. An AI would "learn" that and who could reliably catch it doing so?

von Marwitz

In my gut I firmly believe that the AI cheats. At least in Civilization. If AI or a biological opponent thinks they have to cheat to beat me I take it as a compliment.

 

[bendizoid]

The AI of Twilight Struggle cheats to.

 

[Old Noob]

PresterJohn: I think I saw a cousin of your version in Portland OR 9/23/24.
Was doing my annual Powells Books visit (along with a couple CD stores) and spotted it.

 

[crocodile_john]

Computers are good at playing chess because chess is principally a game of calculation. If a computer was powerful enough to play ASL do you think it could be good at it? Is ASL also a game of calculation or does some sort of artistic or intuitive skill come into it?

I think even if the rules were completely codified that would not be enough, the computer could make many legal moves but never win. It would be hard to implement a 'winning' strategy.
The alternative would be to use AI to learn by 'training' on previous games.
However, this highlights another problem... there would need to be a way to 'record' ASL games so a computer could analyse them. Some kind of formalized logical notation for a generic ASL move/turn would be needed..This must be achievable somehow.. in the same way chess games are easily recorded using the chess notation....
I guess vasl does this to a degree in a vasl log, but not sure how that works wrt sequential state.
It would be great if a game could be recorded manually (with a pen and paper, move by move) which could be subsequently captured as a simple text file... a computer could definitely learn to play by analysing many, many of those files for a specific scenario.

 

[tommyl]

It would be great if a game could be recorded manually (with a pen and paper, move by move) which could be subsequently captured as a simple text file

Sounds like the Series Replays in the early Annuals.

 

[pensatl1962]

We need to get Jeff Bezos and Elon Musk to play ASL...

von Marwitz

Fixed it: We need to get Jeff Bezos to play ASL...

Keep Elon Musk far the hell away from our hobby.