AI "overperformed"

[Lidozin]

I've seen many claims here suggesting that the AI (aircraft type) possesses "unrealistic" or "supernatural" flight physics. However, since all the aerodynamic and performance data used by the AI are available in the .lua file, it's possible to compute its actual characteristics and compare them directly with those from the real aircraft’s technical documentation.

Let’s start with the flight envelope. The original manual doesn’t provide a complete envelope chart, so we'll have to compare individual performance points mentioned throughout the text. When these reference points are plotted over the AI’s calculated envelope, the result speaks for itself. At the very least, in this aspect, there’s no evidence of any "supernatural" behavior — the AI’s performance stays well within the bounds of what's expected based on real-world data.

Edited Tuesday at 11:32 AM by Lidozin
Missed text

[AndyJWest]

Now plot sustained turn speeds and rates.

[Lidozin]

The next parameter to be evaluated is rate of climb. Unlike with the flight envelope, the documentation does include charts for both maximum rate of climb and the indicated airspeed at which it is achieved, across different altitudes.

By plotting the AI aircraft’s rate-of-climb versus IAS at various altitudes, we can determine both the peak climb rate and the corresponding airspeed at three key points: sea level, 5,000 meters, and 10,000 meters.

Once again, we observe a very close match between the AI’s energy performance and the documented data — strong evidence that the flight model behaves realistically in this regard as well.
All calculations were performed for a nominal aircraft weight of 5,000 kg.

Edited Tuesday at 11:56 AM by Lidozin

[Lidozin]

For turn time at 1,000 meters altitude, the documentation provides calculated reference data. This point has been plotted on the graph of the AI aircraft’s computed turn time.

To keep the analysis concise, we'll limit the comparison to two altitude points — 1,000 meters and 10,000 meters.
For 10,000 meters, the documentation includes a turn performance chart obtained from actual flight tests; those reference points will be overlaid on the AI's calculated curve.

null

 

null

Edited Tuesday at 12:18 PM by Lidozin

[Lidozin]

Thus, the casual claims that the AI aircraft MiG-15 possesses "supernatural" power can be put to rest. The core energy-related parameters — both in 1g flight (such as maximum speed and specific excess power) and in sustained turns at zero excess power — show very strong agreement with available reference data.

Edited Tuesday at 12:28 PM by Lidozin

[Hiob]

The main problem with AI aircraft in dogfights is that they don't seem to bleed much energy when pulling tight or at least regain those energies very quickly.
i don't think the problem is mainly fixed parameters like climb rate or sustained turn rate.
Or in other words, where the player has to constantly balance his energy budget, the AI pilot doesn't seem to care much.

But - at least from my part - that is just subjective observation. Didn't do any "scientific" research on the matter.

[Raven (Elysian Angel)]

3 minutes ago, Hiob said:

The main problem with AI aircraft in dogfights is...

On top of any manoeuvring characteristics, there's the problem of their omniscient situational awareness through vision that's unobstructed by canopy frames and aircraft structures, and vastly overperforming sensor abilities in general. This allows the AI to perform their manoeuvres as perfectly as their FM allows since their decision making will always be perfect. 

[Hiob]

Just now, Raven (Elysian Angel) said:

On top of any manoeuvring characteristics, there's the problem of their omniscient situational awareness through vision that's unobstructed by canopy frames and aircraft structures, and vastly overperforming sensor abilities in general. This allows the AI to perform their manoeuvres as perfectly as their FM allows since their decision making will always be perfect. 

very true

[Lidozin]

2 hours ago, Raven (Elysian Angel) said:

On top of any manoeuvring characteristics, there's the problem of their omniscient situational awareness through vision that's unobstructed by canopy frames and aircraft structures, and vastly overperforming sensor abilities in general. This allows the AI to perform their manoeuvres as perfectly as their FM allows since their decision making will always be perfect. 

This is easy to verify with a simple test setup: position yourself relative to the AI aircraft so that you're firmly within its rear blind cone — specifically, low-six o'clock, about 100–200 meters behind and below, matching its speed to maintain position.
Maintain pursuit while staying within that no-visibility sector.

If the AI still detects you and reacts — despite following its flight plan and having orders to engage the first contact it sees — then it truly does have 360-degree situational awareness, which would be unrealistic.

Next, briefly move just outside the blind cone — into a zone where a real pilot could reasonably acquire a contact visually — and observe the difference in reaction time and behavior.

Edited Tuesday at 03:46 PM by Lidozin

[Lidozin]

2 hours ago, Hiob said:

The main problem with AI aircraft in dogfights is that they don't seem to bleed much energy when pulling tight or at least regain those energies very quickly.
i don't think the problem is mainly fixed parameters like climb rate or sustained turn rate.
Or in other words, where the player has to constantly balance his energy budget, the AI pilot doesn't seem to care much.

But - at least from my part - that is just subjective observation. Didn't do any "scientific" research on the matter.

If the induced drag component of the aircraft's L/D polar is known (and the polar itself is available from wind tunnel tests up to high values of CL, and the model's energy performance matches the real aircraft both at low lift coefficients (e.g. specific excess power in 1g flight) and at high CL (e.g. sustained turn at zero excess power), then it follows that energy loss during high-load, non-sustained maneuvers — where CL is even higher — will also closely match the real aircraft.

I find it hard to believe that whoever tuned this flight model would have ignored such a rich dataset — especially given how thoroughly the design bureau compiled and published the aircraft’s aerodynamic characteristics in the official technical documentation.
nullnull

[Hiob]

5 minutes ago, Lidozin said:

If the induced drag component of the aircraft's L/D polar is known (and the polar itself is available from wind tunnel tests up to high values of CL, and the model's energy performance matches the real aircraft both at low lift coefficients (e.g. specific excess power in 1g flight) and at high CL (e.g. sustained turn at zero excess power), then it follows that energy loss during high-load, non-sustained maneuvers — where CL is even higher — will also closely match the real aircraft.

I find it hard to believe that whoever tuned this flight model would have ignored such a rich dataset — especially given how thoroughly the design bureau compiled and published the aircraft’s aerodynamic characteristics in the official technical documentation.
nullnull

I happen to think, that it isn't for the most part a limitation of available data, but of computing power. If any AI aircrafts flight dynamics would be thoroughly computet in realtime, considering all necessary parameters, it would most likely melt your CPU. Therefore they rely basically on lookup tables with limited data points.
I may be completely off here, I'm just speculating. But it makes sense to me.

That's not meant to be an excuse or something. I would like the best possible (realistic) flight model and capabilities for AI as much as the next guy.

[Raven (Elysian Angel)]

If the AI's FM was that good, there would be no need for the new GFM.

[Lidozin]

null

17 minutes ago, Hiob said:

I happen to think, that it isn't for the most part a limitation of available data, but of computing power. If any AI aircrafts flight dynamics would be thoroughly computet in realtime, considering all necessary parameters, it would most likely melt your CPU. Therefore they rely basically on lookup tables with limited data points.
I may be completely off here, I'm just speculating. But it makes sense to me.

That's not meant to be an excuse or something. I would like the best possible (realistic) flight model and capabilities for AI as much as the next guy.

You are fundamentally mistaken here. When we talk about AI aircraft — or even more advanced flight models — in the context of maneuvering characteristics (i.e. the ability to change trajectory), we're primarily discussing the motion of the aircraft in space as a point mass.

Simulating this type of motion has been possible since long ago — even a ZX Spectrum or a PC XT was sufficient for solving such equations. And today, it's entirely feasible to simulate thousands of aircraft using an AI-level flight model in real time.

A few lookup tables — lift and drag coefficients, thrust vs. altitude and speed — along with some auxiliary data, and voila: you have a highly accurate trajectory and maneuvering model.

[Lidozin]

19 minutes ago, Raven (Elysian Angel) said:

If the AI's FM was that good, there would be no need for the new GFM.

As far as I understand from press releases, the advantage of GFM is that it provides a much more natural simulation of aircraft behavior during short-period motion — that is, rotation around the center of mass — while preserving the accuracy of the older, trajectory-based model.
The very same accuracy we’ve been seeing demonstrated throughout this discussion thread.

And that’s exactly why we’re looking forward to the new model with such anticipation.

https://www.digitalcombatsimulator.com/en/news/2021-12-03/

Edited Tuesday at 04:48 PM by Lidozin

[Hiob]

You can‘t just put player flight models and AI flight models in one bucket. The first is only computed once the latter potentially dozens of times simultaneously. 
and I would reject the notion that the AI can be simplified to a point mass. Things like AOA (to pick but ONE prominent example are strongly correlated to the loss of energy e.g.

[Schmidtfire]

You can show numbers and charts all day long, but the outcome within DCS World is all that matters.

AI SFM numbers needs to be offset from reality in order to create something that is closer to "real" performance.

DCS developers (in my opinion) sometimes has a tendency to put the diagrams above everything else...

...and everything is good if the math is correct -  even if the final result is far from reality. 

 

[Lidozin]

2 hours ago, Hiob said:

You can‘t just put player flight models and AI flight models in one bucket. The first is only computed once the latter potentially dozens of times simultaneously. 
and I would reject the notion that the AI can be simplified to a point mass. Things like AOA (to pick but ONE prominent example are strongly correlated to the loss of energy e.g.

Let me clarify a few points, since there seems to be a misunderstanding about what trajectory-based (or “point mass”) flight models actually include.

First of all, the notion that a trajectory model ignores angle of attack is simply incorrect. In fact, angle of attack (AoA) is one of the key input parameters used to define the aerodynamic polar — the lift and drag coefficients are tabulated precisely as functions of AoA. When a simulation computes the net force on the aircraft at any moment, it determines AoA from the velocity vector and attitude, then looks up the corresponding L and D values from that polar. This is standard in both AI models and performance simulation tools used in real-world aerospace.

So yes — energy loss due to drag is inherently tied to AoA in such models. It’s not being ignored; it’s the foundation for how excess power, turn rate, and climb performance are derived.

Spoiler

Now, what the full (6-DOF) flight model does on top of that is compute the full system of differential equations that governs rotational dynamics — for example, to model how the aircraft transitions to a given angle of attack: how long it takes to get there, whether there’s overshoot, whether the response is stable or divergent. It also integrates equations of motion around all axes to resolve angular velocities and orientations over time.

By contrast, the trajectory-based model uses simplified algorithms to estimate AoA, bank, and sideslip — based on typical solutions to those same equations, assuming the aircraft is in a quasi-steady, coordinated state. That’s a deliberate and well-understood simplification, and for many purposes — including air combat simulation — it’s more than sufficient.

As for the computational load argument: yes, AI models are simpler than full 6-DOF player models, but that doesn’t mean they’re “dumb.” It just means their short-periodic behavior is driven by precomputed aerodynamic data, rather than real-time airframe dynamics simulation. This approach has been successfully used in industry-level simulations of air combat scenarios, and it allows very accurate modeling of maneuver performance — as long as the aerodynamic inputs (like the polar and thrust tables) are accurate.

So respectfully, I think you're overestimating what “not being a 6-DOF model” implies. The point-mass model isn’t a toy — it’s a well-established engineering approach that, when properly implemented, produces excellent agreement with real aircraft energy behavior, as shown above.

And just to be clear: the graphs shown earlier aren’t “just graphs” — they are, in fact, the results of a flight dynamics simulation in their own right. They were derived analytically using publicly available aircraft data, processed through well-known aerodynamic and energy modeling methods. In other words, what you're seeing is essentially the same type of output you’d get from a real-time simulation — just computed offline using standard techniques.

 

[Lidozin]

1 hour ago, Schmidtfire said:

You can show numbers and charts all day long, but the outcome within DCS World is all that matters.

AI SFM numbers needs to be offset from reality in order to create something that is closer to "real" performance.

DCS developers (in my opinion) sometimes has a tendency to put the diagrams above everything else...

...and everything is good if the math is correct -  even if the final result is far from reality. 

 

I had absolutely no intention of defending anyone — neither dissatisfied users, nor, certainly, the developers.
My goal is to establish the facts, based on knowledge — as lofty as that may sound.

If this model had shown significant deviations from the reference documentation, I would have gladly published those findings.

This particular aircraft caught my attention because it's one of the most frequently criticized by users, and at the same time, one of the best-documented — both in terms of source data, which undoubtedly gave the developers a solid foundation to work with, and in terms of flight performance, which provides a solid basis for comparing the model with the real aircraft in detail.

[Pikey]

This was the same type of reasoning provided by ED, that their model is using real world data so the results are perfect.

I can and do edit SFM's and i'm doing some right now and it really doesnt matter what you put in to the model, it relies on the model being perfect, not the data.

You tweak it to behave in the game like something that works well. All of a sudden you broken it somewhere else. Same model, have to change the real world data to fit.

 

[Lidozin]

2 hours ago, Pikey said:

This was the same type of reasoning provided by ED, that their model is using real world data so the results are perfect.

I can and do edit SFM's and i'm doing some right now and it really doesnt matter what you put in to the model, it relies on the model being perfect, not the data.

You tweak it to behave in the game like something that works well. All of a sudden you broken it somewhere else. Same model, have to change the real world data to fit.

 

I’d like to respond to your points, because I think there’s a fundamental misunderstanding here — not about opinions, but about how flight models actually work.

First of all, a trajectory-based model (point mass model) is not something that can be "good" or "bad" in itself. It's simply a set of well-known, well-defined differential equations that describe the motion of an aircraft CoG under the influence of aerodynamic, thrust, and gravity forces. Solving these equations — through numerical integration — gives us the actual flight trajectory, including climb, acceleration, turn performance, and so on.

Spoiler

What determines whether such a model is accurate or not is not whether it "feels right" in a game, but whether it reproduces known flight characteristics of the real aircraft. These include:

• the flight envelope (available from test data),

• maximum climb rates and the speeds at which they occur,

• sustained and instantaneous turn rates,

• energy performance (specific excess power),

• and so on.

All of these characteristics are known and documented — especially for legacy aircraft — and are the benchmark against which a simulation model is validated.

You mentioned that “you tweak it to behave in-game like something that works well,” but that’s exactly the problem: "works well" is subjective unless it's backed by reference data. If you're modifying parameters by feel, to make the model "fly right," you're essentially distorting the data to compensate for a misunderstanding of the model structure.

In fact, the graphs shown earlier in this thread are not just visualizations — they represent the analytical outcome of the same trajectory equations, using real aerodynamic input data. They describe how the aircraft behaves across different regimes: acceleration, climb, turn performance, etc. And when the simulated data closely matches the known performance of the real aircraft, the model is working — regardless of whether it was coded as SFM, EFM, or otherwise.

Lastly, it’s worth emphasizing that trajectory models — even fairly simple ones — have been used for decades in serious aerospace simulation. What matters is the quality of the input data and the validation process, not whether the model includes every rotational dynamic detail. That’s what 6-DOF models are for — and even then, they too rely on accurate aerodynamic input to produce valid behavior.

 

[Pikey]

On 7/8/2025 at 11:39 PM, Lidozin said:

I’d like to respond to your points, because I think there’s a fundamental misunderstanding here — not about opinions, but about how flight models actually work.

First of all, a trajectory-based model (point mass model) is not something that can be "good" or "bad" in itself. It's simply a set of well-known, well-defined differential equations that describe the motion of an aircraft CoG under the influence of aerodynamic, thrust, and gravity forces. Solving these equations — through numerical integration — gives us the actual flight trajectory, including climb, acceleration, turn performance, and so on.

  Hide contents

What determines whether such a model is accurate or not is not whether it "feels right" in a game, but whether it reproduces known flight characteristics of the real aircraft. These include:

• the flight envelope (available from test data),

• maximum climb rates and the speeds at which they occur,

• sustained and instantaneous turn rates,

• energy performance (specific excess power),

• and so on.

All of these characteristics are known and documented — especially for legacy aircraft — and are the benchmark against which a simulation model is validated.

You mentioned that “you tweak it to behave in-game like something that works well,” but that’s exactly the problem: "works well" is subjective unless it's backed by reference data. If you're modifying parameters by feel, to make the model "fly right," you're essentially distorting the data to compensate for a misunderstanding of the model structure.

In fact, the graphs shown earlier in this thread are not just visualizations — they represent the analytical outcome of the same trajectory equations, using real aerodynamic input data. They describe how the aircraft behaves across different regimes: acceleration, climb, turn performance, etc. And when the simulated data closely matches the known performance of the real aircraft, the model is working — regardless of whether it was coded as SFM, EFM, or otherwise.

Lastly, it’s worth emphasizing that trajectory models — even fairly simple ones — have been used for decades in serious aerospace simulation. What matters is the quality of the input data and the validation process, not whether the model includes every rotational dynamic detail. That’s what 6-DOF models are for — and even then, they too rely on accurate aerodynamic input to produce valid behavior.

 

Ah. "Yoyo mk2". "it's right because it is physics and that is unquestionable". Argument over. You based your argument on data which is physics so it's unquestionable. 

 

Ive got some news for you. You built your premise on an assumption, that the physics model is being used all the time.

See those routines AI performs... See how they follow patterns...see how they snap together so neatly... or maybe you aren't looking.

Let me tell you something you apparently don't know or have ever seen. it will come as a shock.

 

Ai doesn't use physical models during all aspects of flight. 

 

Have you even watched AI forming up and suddenly braking in the air 50 knots suddenly stopping like a car crash?

Have you seen AI brake on the runway? 

Have you seen the points of the SFM curve when the AI gets stuck between two drag coefficients and snaps between them causing it to flip and jerk.

Have you seen the AI warbirds flying around with no engine floating at cruising speed, not losing altitude?

Have you seen the AI in this game perform something that is not physically possible, despite apparently using a model that is physically sound?

I think you don't play this game!

You don't need to be Sir Isaac Newton to know an apple falls down, it doesn't float. Physics is not the problem here. You are arguing about physics when the topic is software, where magic is possible and under that illusion is this game. You talk models but have assumed it's being used, at least all the time. I'm here to tell you they aren't, much like the people who think little 'mitochondria' come out of the nose of aircraft with radar...they don't either, it's software.

 

I don't want to hear about models that are not applying to this simulation, it is a worthless and impractical spend of my time when every year I report some strange behaviour where it might be the most accurate nasa supercomputer model using quantum mechanics for all I care but if a plane is floating upside down or even if it's just got the attributes of a rocket or a snail, no one cares, it's wrong, end of conversation.

 

 

[LeCuvier]

I totally agree with @Pikey in every point he makes. 

 

[Dragon1-1]

On 7/8/2025 at 9:50 PM, Pikey said:

This was the same type of reasoning provided by ED, that their model is using real world data so the results are perfect.

And that's the problem - they are perfect. AI is consistently achieves and maintains the aircraft's performance limits in ways a human could never manage. This was noted, for instance, by Reflected in his climbing with AI warbirds video. If you're in perfect trim, have your engine settings just right, and don't make any mistakes, you can keep up with the AI. I'm pretty sure no WWII pilot could actually set up his aircraft this well (at least that got improved, but only by capping the AI engine power). A dogfight is ultimately decided by who makes the fewest mistakes, and the way AI flies, small errors inherent to every human's piloting simply aren't there.

On 7/9/2025 at 12:39 AM, Lidozin said:

I’d like to respond to your points, because I think there’s a fundamental misunderstanding here — not about opinions, but about how flight models actually work.

The fundamental problem with that model is, it doesn't tell the whole story. As mentioned above, it's dry maths, not accounting for how the plane is actually flown. This, BTW, is why modern jets suffer from this issue much less than vintage ones (sure enough, the AI FM has been designed for modern jets). Nobody flies the MiG-15 on the numbers all the time, but in the F-16, the computer is doing much of the flying for you. If you want to get the best turn rate, just haul back on the stick and presto. This will get you in trouble in vintage planes, but AI can ride the exact point between too hard and too little pull.

[MAXsenna]

Thanks @Pikey! Very well written!

Sent from my SM-A536B using Tapatalk

[Lidozin]

11 hours ago, Pikey said:

Ah. "Yoyo mk2". "it's right because it is physics and that is unquestionable". Argument over. You based your argument on data which is physics so it's unquestionable. 

 

Ive got some news for you. You built your premise on an assumption, that the physics model is being used all the time.

See those routines AI performs... See how they follow patterns...see how they snap together so neatly... or maybe you aren't looking.

Let me tell you something you apparently don't know or have ever seen. it will come as a shock.

 

Ai doesn't use physical models during all aspects of flight. 

 

Have you even watched AI forming up and suddenly braking in the air 50 knots suddenly stopping like a car crash?

Have you seen AI brake on the runway? 

Have you seen the points of the SFM curve when the AI gets stuck between two drag coefficients and snaps between them causing it to flip and jerk.

Have you seen the AI warbirds flying around with no engine floating at cruising speed, not losing altitude?

Have you seen the AI in this game perform something that is not physically possible, despite apparently using a model that is physically sound?

I think you don't play this game!

You don't need to be Sir Isaac Newton to know an apple falls down, it doesn't float. Physics is not the problem here. You are arguing about physics when the topic is software, where magic is possible and under that illusion is this game. You talk models but have assumed it's being used, at least all the time. I'm here to tell you they aren't, much like the people who think little 'mitochondria' come out of the nose of aircraft with radar...they don't either, it's software.

 

I don't want to hear about models that are not applying to this simulation, it is a worthless and impractical spend of my time when every year I report some strange behaviour where it might be the most accurate nasa supercomputer model using quantum mechanics for all I care but if a plane is floating upside down or even if it's just got the attributes of a rocket or a snail, no one cares, it's wrong, end of conversation.

 

 

You’ve raised a number of observations about strange or buggy AI behavior — some potentially valid, others harder to verify — but I think we may be talking past each other.

The original discussion wasn’t about general AI behavior across all mission stages. It was specifically about how the AI performs in a dogfight, and whether the AI's flight model during combat is based on real aerodynamic parameters that correspond to those of the real aircraft.

So let me ask directly:
What exactly, in your experience, seems unrealistic or broken in AI behavior during a 1-on-1 dogfight — with both human and AI flying the same aircraft, from the same initial conditions?

If there’s a mismatch in energy performance, turn rate, climb, etc., under those circumstances — that’s something worth looking into. But if the concerns are about form-up logic, taxiing behavior, or scripted transitions, those are separate layers of the simulation, and not what’s being discussed when we refer to the AI using a physics-based trajectory model during combat.

Let’s isolate the question to air combat maneuvering performance. That’s the only way to make progress on whether the model is being applied correctly — or not — in that context.

Personally, I tend to focus on this specific aircraft and enjoy 1-on-1 dogfights in matching types, precisely because they allow a fair comparison of skill and energy management. I don’t spend much time observing AI in other scenarios, so I leave those potential bugs to others — for me, the duels are more than enough.