Chat GPT and Wargame AI

[Elxaime]

Hi All -

With all the fuss over Chat GPT, some have asked how this development can relate to video gaming.  For example, in online roleplaying worlds, AI is now being studied to make non-player character (NPC) interactions more varied and real seeming.

I wonder if this same technology can create better AI opponents for wargames like Panzer Campaigns?  I am not versed in the tech, but if the idea is that Chat GPT, if given properly focused parameters, can mimic human responses and decision making, can't this be "plugged in" to wargaming to give us better AI opponents?

Maybe this is being fanciful, but could such AI not only model generic AI generals, but actually model specific ones.  Imagine an AI that thinks like Wellington opposite you in Waterloo.

Thoughts?

JVJ

[A Canadian Cat]

I wonder if this same technology can create better AI opponents for wargames like Panzer Campaigns?  I am not versed in the tech, but if the idea is that Chat GPT,

The short answer is "no". Chat GPT is a language processor and it is very good at interpreting questions and writing answers. It is not a tactical thinker. In fact it is not thinking at all really.

if given properly focused parameters, can mimic human responses and decision making, can't this be "plugged in" to wargaming to give us better AI opponents?

Having said what Chat GPT can and cannot do it is absolutely possible to train a machine learning system to do tactics. What you would need would be a large variety of scenarios, units and terrain and then record the battles run by many many humans over and over and then let the AI learn how they behaved in the various situations with a variety of resources. Then the model could be used to make tactical decisions in a new novel scenario.

With enough quality examples it would be quite formidable. With the wrong data it would suck. For example if you recorded novice players fighting against the current AI the result would likely not be very good. Humans often do things against the AI they only do because they know how the AI works. Choices like that don't usually work against a human opponent. If you recorded many thousand battles between experienced human opponents there is a good chance novices would have a very hard time against that AI

[Ricky B]

One item to note in regards to building a strong AI using examples is that it needs to be based on a bit more of a "memory" model of what has happened in prior turns. I have noted over the years that many game publishers brag that the AI they built has no more information than a human player. However, in all the cases I am aware of, they don't build a memory, so there is no modeling of a plan in a wargame, at least, based on what a player has seen since the start of the game.

So out of sight in the current turn means out of mind in deciding on a plan. Having no FOW for the AI helps a bit and of course gives too much information, but even that isn't solution as it isn't going to help as it would for a player remembering that over the past 5 turns the opponent has shifted the 10th Tank Corps for the A axis over to the C axis, so this shift needs to be considered too.

So a training plan would help (a lot!) but it would have to capture a lot more than just the current turn to really determine a plan to follow.

[Elxaime]

Maybe I misunderstand what AI/machine learning can and cannot do. But what I was envisioning was something like how they expect AI to tackle coding itself. Initially, you would ask the AI to learn off thousands and thousands of AI v. AI games (eliminating the issue of finding humans to play humans to generate data). The current WDS AI for example, which we can call WDS 1.0. It would then select for the most successful approaches taken by WDS 1.0 and use that data to develop WDS 2.0, the "next generation" of basic AI. WDS 2.0 would then be pitted against itself to develop the next generation of raw data, which would be the foundation for coding WDS 3.0, etc. Human coders can intervene along the way to guide evolution, but the raw generation of data to learn from and the coding of the next generation of test-bed AI would be automated. Or something, heh.

[A Canadian Cat]

Maybe I misunderstand what AI/machine learning can and cannot do.  But what I was envisioning was something like how they expect AI to tackle coding itself.  Initially, you would ask the AI to learn off thousands and thousands of AI v. AI games (eliminating the issue of finding humans to play humans to generate data).

True that is another way to do it. That's how the Go and Chess AIs were programmed. Actually that would be a better choice probably. The issue is there is commercially available modeling / training AI packages for what I was suggesting earlier. I am not sure if the self learning systems used to play Go and Chess are actually something that a commercial company can buy for their own purposes. I don't know enough about the AI market to answer that.

[Scud]

I asked ChatGPT if it could mine some bitcoin for me. As expected the answer was no. It explained why, but at that point, I no longer cared. 

[Elxaime]

Another interesting article on generative AI and wargaming:

It Is Time to Democratize Wargaming Using Generative AI (csis.org)

This nugget at the end caught my eye:

To that end, the Department of Defense needs to accelerate its support for efforts like TF LIMA—the new generative AI task force—and experiments like the Global Information Dominance Exercise. More importantly, services need to start funding copilots and other unclassified AI tests at lower echelons while studying how best to train military professionals to work with—not against—models that aggregate data. In all likelihood, this movement will require significant changes to professional military education to include practicums on data science, statistics, research methods, and red teaming.

Could not the game designers get contracts and use the work they do for the government to improve the game AI's for our wargames?

[parmenio]

Could not the game designers get contracts and use the work they do for the government to improve the game AI's for our wargames?

The problem is if we're all working on government contracts, we're not going to be working on anything else. The other potential issue is who ends up with the IPR for this nice shiny new AI that they've paid for. 

[parmenio]

Maybe I misunderstand what AI/machine learning can and cannot do.  But what I was envisioning was something like how they expect AI to tackle coding itself.  Initially, you would ask the AI to learn off thousands and thousands of AI v. AI games (eliminating the issue of finding humans to play humans to generate data).

True that is another way to do it. That's how the Go and Chess AIs were programmed. Actually that would be a better choice probably. The issue is there is commercially available modeling / training AI packages for what I was suggesting earlier. I am not sure if the self learning systems used to play Go and Chess are actually something that a commercial company can buy for their own purposes. I don't know enough about the AI market to answer that.

Stockfish which is one of the top Chess Engines in the world is actually Open Source:

https://stockfishchess.org/

This is an interesting YouTube video on how Stockfish works (one of the Stockfish developers also appears):

Why AI Chess Bots Are Virtually Unbeatable (ft. GothamChess) | WIRED (youtube.com)

AlphaZero (another top Chess Engine) was (apparently) programmed with the rules of chess and then played millions of games against itself and was then able to beat anyone and anything at Chess, Go and Shogi:

AlphaZero: Shedding new light on chess, shogi, and Go - Google DeepMind

I think one of "yeah buts" against AlphaZero when it beat Stockfish, for example, is that its Neural Network was created by using Google's server farm whereas Stockfish runs on a laptop (that might well be an exaggeration of the extremes but you get the point). AlphaZero is definitely not Open Source.

Certainly AI vs. AI would be an approach but it can difficult to equate Chess AI to Wargame AI. Take the board - chess is effectively played on a map of 8 hexes by 8 hexes with 6 unit types (and it's only ever that - no "what if" scenarios here). There is no terrain, no elevation, no FOW. There is  also no randomisation. If a piece can be attacked and is it's eliminated and there are no die rolls involved in that calculation. What 'm getting at is that all of this doesn't make the "machine learning" aspect simple.

Neural networks would likely be the way to go and there are certainly Open Source options there. But as ever all of this would takes a huge amount of time and needs deepish pockets to fund the work. It's effectively R&D with no guarantees of timely results.

[Elxaime]

Maybe I misunderstand what AI/machine learning can and cannot do.  But what I was envisioning was something like how they expect AI to tackle coding itself.  Initially, you would ask the AI to learn off thousands and thousands of AI v. AI games (eliminating the issue of finding humans to play humans to generate data).

True that is another way to do it. That's how the Go and Chess AIs were programmed. Actually that would be a better choice probably. The issue is there is commercially available modeling / training AI packages for what I was suggesting earlier. I am not sure if the self learning systems used to play Go and Chess are actually something that a commercial company can buy for their own purposes. I don't know enough about the AI market to answer that.

Stockfish which is one of the top Chess Engines in the world is actually Open Source:

https://stockfishchess.org/

This is an interesting YouTube video on how Stockfish works (one of the Stockfish developers also appears):

Why AI Chess Bots Are Virtually Unbeatable (ft. GothamChess) | WIRED (youtube.com)

AlphaZero (another top Chess Engine) was (apparently) programmed with the rules of chess and then played millions of games against itself and was then able to beat anyone and anything at Chess, Go and Shogi:

AlphaZero: Shedding new light on chess, shogi, and Go - Google DeepMind

I think one of "yeah buts" against AlphaZero when it beat Stockfish, for example, is that its Neural Network was created by using Google's server farm whereas Stockfish runs on a laptop (that might well be an exaggeration of the extremes but you get the point). AlphaZero is definitely not Open Source.

Certainly AI vs. AI would be an approach but it can difficult to equate Chess AI to Wargame AI. Take the board - chess is effectively played on a map of 8 hexes by 8 hexes with 6 unit types (and it's only ever that - no "what if" scenarios here). There is no terrain, no elevation, no FOW. There is  also no randomisation. If a piece can be attacked and is it's eliminated and there are no die rolls involved in that calculation. What 'm getting at is that all of this doesn't make the "machine learning" aspect simple.

Neural networks would likely be the way to go and there are certainly Open Source options there. But as ever all of this would takes a huge amount of time and needs deepish pockets to fund the work. It's effectively R&D with no guarantees of timely results.

I wonder if the time/deep pockets aspects might be mitigated by some sort of contract with the US military or another branch of the US Government (which I understand in the past has contracted with some wargame developers - I seem to recall John Tiller might have had some of that work - perhaps I am thinking of someone else).  Then, the model you developed gets improved by not only competing AI v AI, but by allowing the public to play against the AI in a free to play online game setting.