Deep learning AI pilots

[Motomouse]

Computers learned to play GO with deep learning.

I am sure they can play DCS with deep learning as well.

Dont integrate the AI into DCS. Let the AI learn to play the game aka fly the planes as clients.

I know ED has not the resources like DeepMind or OpenAI, but I am sure as well there are entities that have the resources and operate something like this already

[Insonia]

there is nothing new in this idea. There is just no commercial value in ED making this happen.

DCS has very sophisticated export and import functions. from sensor data to imagery, etc.. and accept input through multiple channels.

You learn a little bit of Python and use existing machine learning packages/models. then you can start train it today.

Someone has already tried some interesting things:

However, it was quite different to make it an actual working, economically viable, and profitable product. 

 

If there is commercial value, they could just outsource it to professional machine learning company, would be faster and cheaper; they wouldn't need to risk their money and manpower on something not in their expertise. The current existing ED game system is enough to feed the training cycle.

And as long as we have enough players like you willing to pay. We may have a chance. 

Edited October 25, 2023 by Insonia

[Dragon1-1]

This work will not help our AI fly better. Why not? Simple, the AI doesn't use the same FM humans do. The other WWII sim does take that approach, but the performance cost is steep. They still don't have four engine bombers because of this. The experiment above basically substituted the neural network based controller for the player, this would not work with regular kind of AI that we have in DCS.

What we need is GFM, which will give AI a more lifelike, but less performance intensive FM. AI decision logic is also being worked on, but in most cases, it's sufficient. It would be possible to make the AI, even now, follow the rules of formation flying, because those rules are fairly simple and very deterministic. Rules that are easy to follow make for safe flying IRL, so making DCS AI just stick to the book exactly does not require any fancy tech. Robots are good at following the rules.

[Motomouse]

The key point is that an external AI won't use the AI flight model, and to DCS, it will be indistinguishable from a human pilot, utilizing the same flight model as we do 

You do not "program" the AI to fly an artifical FM. The AI learns to fly DCS planes like humans do. That is what deep learning is about.

The track files of DCS of the human players provide nice traininng data for the neural network. I am sure somebody is doing this already

Edited October 25, 2023 by Motomouse

[cfrag]

11 hours ago, Motomouse said:

That is what deep learning is about.

The point still stands what ED's business interest would be for what currently seems to be a significant investment. What do they get out of it, how can they recoup that cost and make more on top? Currently we have 

1. Play DCS

2. Add Deep Learning AI 

3. ???

4. Profit

We need more clarity on 3 even if we all agree on 2.

[Motomouse]

@cfrag: Why that fixation on ED. Because I noted the wish on the DCS wishlist? 

... Welcome to the Wish List section for the core game, this is for all those things you would love to see added to DCS World. This is a safe zone for free expression of ideas, there are no bad ideas that doesn't mean every idea will be added or can be added. ...

Edited October 26, 2023 by Motomouse

[Dragon1-1]

18 hours ago, Motomouse said:

The key point is that an external AI won't use the AI flight model, and to DCS, it will be indistinguishable from a human pilot, utilizing the same flight model as we do 

Not feasible. What you want to do is make every mission an MP mission, with possibly dozens of ML-enabled pilots... running on a single computer? Sorry, that's not happening, because nobody has a PC that could handle this, plain and simple.

This alone places your idea firmly in "pie in the sky dream" category. If you meant standalone AI client machines to fill empty slots on a multiplayer server, that could compete with human players... then perhaps it could be developed a standalone product server owners could buy. If you mean bog standard AI tankers and such for singleplayer, the idea is as stupid as it is impossible (especially given that most of those aircraft don't have a playable FM).

6 hours ago, cfrag said:

The point still stands what ED's business interest would be for what currently seems to be a significant investment.

Replace "significant investment" with "practical impossibility". The arrangement proposed does not work. Period. The only way something like this could work is as stated above, a product, likely steeply priced, intended for server owners to add human replacement bots. And even then, you run into the question of whether it's OK to use community tracks to train a commercial product (of course, it'd also learn as it is used).

[Motomouse]

I enjoy playing go versus ai on my gpu and ai in the cloud.

I am sure I would enjoy dogfighting one on one against ai as well.

Some of the competent go ai were developed open source with an managable effort by a small group of people after the example of deep mind.

Flying a plane is a much easier task than playing go or understanding and producing natural language.

But I do not want to win an argument here. Nobody has to agree with my wishes. Lets wait and see.

[cfrag]

5 hours ago, Motomouse said:

Why that fixation on ED

Simply because of the fact that they alone put the AI into DCS

5 hours ago, Motomouse said:

there are no bad ideas that doesn't mean every idea will be added or can be

I didn’t say the idea is bad - I too would love to see Deep Learning supplemented AI; not only for enemies, but other stuff like ATC. I’m merely remarking that expressing wishes is one thing. Wishes that we can make a good case for, on the other hand, may even have a sliver of a chance to be implemented. 

Edited October 26, 2023 by cfrag

[upyr1]

On 10/25/2023 at 2:18 PM, Motomouse said:

The key point is that an external AI won't use the AI flight model, and to DCS, it will be indistinguishable from a human pilot, utilizing the same flight model as we do 

You do not "program" the AI to fly an artifical FM. The AI learns to fly DCS planes like humans do. That is what deep learning is about.

The track files of DCS of the human players provide nice traininng data for the neural network. I am sure somebody is doing this already

 

This idea would only work for AC that are available as mods, becuse no matter what CPU is running the AI, it can only use an available flight model. So you will need to improve the AI flight model. I'm not flat out shooting down the idea, I'm just saying you need better flight models. Once those are in place I think this could be a good idea. It all comes down to whether or not the neural network does a good job of imitating a human pilot. I think this could be a good way to test future AI models 

[Zius]

On 10/25/2023 at 5:18 PM, Insonia said:

There is just no commercial value in ED making this happen.

Of course there is.

Better AI = more enjoyable single player experience = more players.

Bad AI, that is, AI which does inexplicable things like not dropping bombs when supposed to, or AI which dogfights so poorly that it doesn't pose a challenge, does hurt DCS.

Not everybody wants to join multiplayer servers. And also those have their limitations.

In an ideal world, AI opponents should be indistinguishable from human opponents. If that were to happen, this would be a great asset to any combat flight sim.

[Insonia]

13 hours ago, Zius said:

Of course there is.

Better AI = more enjoyable single player experience = more players.

Bad AI, that is, AI which does inexplicable things like not dropping bombs when supposed to, or AI which dogfights so poorly that it doesn't pose a challenge, does hurt DCS.

Not everybody wants to join multiplayer servers. And also those have their limitations.

In an ideal world, AI opponents should be indistinguishable from human opponents. If that were to happen, this would be a great asset to any combat flight sim.

if you read the op. It's about a ML/DL cloud-based "AI" client joining as a client. Very specific type of implementation. You stretch the topic a little bit.

There are a lot of misconceptions about AI in video games. They are actually designed to be BAD in a specific manner. good enough to be functional, but bad enough to be beatable. Todays AI can already annihilate most humans with ease in a game/simulator. But there is no fun if you can't win. And ML is no miracle to turn "AI" humanlike either. Dont get hyped just because they add ML/DL into their works.
Company ofc gonna do whatever it takes to make their product more appealing and profitable. They are not stupid. They haven't done it or will not do it because the (cost of) technology didn't get their yet. and they may have a better place to spend their resources.

The conventional "ai" follows scripted decision trees. The scale of the tree is limited by human factors, but it is easy to adjust and tweak. Fully under developer control. From multibillion-dollar triple A studios to indie developers, they all do in this way.
ML-generated nodes are difficult to understand and tweak from a human perspective. Because the impracticality No game company and developer actively uses ML as their "AI". Instead, the industry treats ML as a statistical research tool, aiding their normal development circle, not a replacement for those features.

Even the most advanced AI company struggles to turn profit, revenue and operation cost barely break even.
Ed's CEO Nick said before in the interview that they are relying on pumping unfinished products to keep the cash flow; otherwise, there will be no DCS. If he spends a million on something, it better have a return and be profitable. And it's not.
There are many commercial level "AI" can enhance current DCS gameplay already. Like synth3d from Maxar for geospatial and mapping. Synthetic voice and language models make radio communication more alive. But it can cost multiple million dollars a year to operate. Consider how small the DCS player base and spending power, most of those are not economically viable . They will wait until those technologies become much cheaper and easier to access.

Edited October 30, 2023 by Insonia

[ValidDisplayName]

On 10/25/2023 at 11:18 AM, Insonia said:

there is nothing new in this idea. There is just no commercial value in ED making this happen.

DCS has very sophisticated export and import functions. from sensor data to imagery, etc.. and accept input through multiple channels.

You learn a little bit of Python and use existing machine learning packages/models. then you can start train it today.

Someone has already tried some interesting things:

However, it was quite different to make it an actual working, economically viable, and profitable product. 

 

If there is commercial value, they could just outsource it to professional machine learning company, would be faster and cheaper; they wouldn't need to risk their money and manpower on something not in their expertise. The current existing ED game system is enough to feed the training cycle.

And as long as we have enough players like you willing to pay. We may have a chance. 

 

Is it against DCS's terms of service to train an AI on this data? If someone wanted to do a personal project to see if they could get a DCS plane to fly on a neural network, does that violate DCS's terms of service to collect that data? 

Many large 3D modeling websites make it against the TOS to use their models in ML training. 

[Tank50us]

Train AI to beat us players...

There are 6 movies and a TV series that explains exactly why this is a bad idea....