AI "overperformed"

[Lidozin]

2 hours ago, Temetre said:

 

@Lidozin Frankly, this thread tells more about pyschology than any aspect of mechanics and simulation. You wrote these two things in the same post:

First this:

And then this:

In one you say everyone else is wrong, because factual arguments have a hard time against group consensus (which is a funny thing to say btw).

In the second you say you dont need to present facts, analysis or evidence regarding applicability of your analysis, because group consensus supports your position (also funny in context of this topic).

 

How do you rationalize these two contradicting lines of argumentation?

 

It seems you're contrasting well-documented facts — including code excerpts and consistent confirmation by developers and users alike — with subjective impressions, and then calling that "group consensus."

But that's exactly the distinction between an engineering approach and opinion-based discussion: one relies on verifiable data, the other on votes.

The real contradiction in your reply is that you question objective sources while accusing others of being overly confident without sufficient basis.

12 hours ago, Pikey said:

The examples provided prove physics is not observed 100% of the time. Only that. Nothing else.

You can state any physics observation you like, but it doesnt serve as evidence to a question of software.

What makes you so certain that the software is being used properly and consistently when you never wrote it and don't have access to it? 

A model can be correct, physics, can be proved. I'm happy that a model is safely beyond reproach. What I'm not convinced of is that its applied correctly or consistently. You cannot read the software, its going on beyond your eyesight. Software does not observe laws therefore you cannot use physics to prove that software conforms.

Now, you marked yourself as the solution in this thread. I don't care about the arrogance of that, but it's a sign that you don't consider any previous or future argument to be of value. SO, since you are th esolution to your own thread, I think you can dispense with everyone else in the world and go back to single player. It's where you shine.
 

The confidence in how the simulation behaves stems from well-established knowledge of the trajectory model in use. There is ample publicly available information describing the aerodynamic model applied in DCS, including the formulas for thrust, drag, and motion. These are not speculative; they’ve been consistently referenced and verified by many within the community.

As philosophy reminds us, practice is the criterion of truth (Karl Marx). If there is any uncertainty, it can — and should — be addressed empirically. One simple and effective test is to task the AI-controlled aircraft with a maximum-rate climb and record the time at which it reaches each successive 1000-metre altitude increment. This approach avoids the need for specialized tools like TacView, requiring only careful observation and a notepad.

It’s a modest investment of effort, but it yields clear data: either the simulation behaves as predicted by the aerodynamic tables, or it does not — and in either case, we move from speculation to grounded evaluation. The test can also be easily shared and repeated by others, allowing for open verification.

Naturally, for consistency, the AI aircraft should have a mass of precisely 5000 kg, and the atmospheric conditions should correspond to ISA.

[Temetre]

Yeah no question, this entire thread is some weird attempt to troll.

[Lidozin]

23 minutes ago, Temetre said:

Yeah no question, this entire thread is some weird attempt to troll.

If presenting transparent methods, reproducible tests, and referencing documented code is “trolling,” then perhaps we've simply redefined what constructive technical discussion looks like.

Let’s stick to substance.

[MAXsenna]

Yeah no question, this entire thread is some weird attempt to troll.

I called it first!

Sent from my SM-A536B using Tapatalk

If presenting transparent methods, reproducible tests, and referencing documented code is “trolling,” then perhaps we've simply redefined what constructive technical discussion looks like.
Let’s stick to substance.

None of that matters when the AI doesn't behave like that.
You failed to answer my QUESTION though. How long have you actually been playing DCS?

Sent from my SM-A536B using Tapatalk

[Raven (Elysian Angel)]

2 hours ago, Temetre said:

Yeah no question, this entire thread is some weird attempt to troll.

And yet it’s been going on for 3 pages

Don’t mind me by the way, I’m only here for the entertainment

[MAXsenna]

And yet it’s been going on for 3 pages
Don’t mind me by the way, I’m only here for the entertainment

Same! I have never seen such a struggle/attempt to gaslight people before.

Sent from my SM-A536B using Tapatalk

[Lidozin]

I'm just lighting the gas to shed some light on the facts.

Some posts above I was told to “shine alone in single player.” Well, I didn't quite go solo — I took ten minutes to run a simple, reproducible test in DCS.

No TacView, no special tools, just a stopwatch and some altitude readings every 1000 m.

The result? Green markers on the reference chart you’ve all seen before — and yes, they sit right on top of the official climb performance curve. No magic, no mystery, no tweaking — just a clean test under ISA conditions and 5000 kg mass.

Turns out, a bit of curiosity and the absence of laziness go a long way.

Now, perhaps we can move from general feelings and philosophical concerns about “how software works” to concrete, measurable outcomes. The data’s there. The method’s simple. The mission is obvious. Anyone can repeat it.
And if the simulation produced results matching the calculations based on the data file, then there’s only one conclusion to draw: the implementation is consistent with the model.

If the claim is that something’s broken — the burden of proof now lies with those making the claim.

[Temetre]

vor 2 Stunden schrieb Raven (Elysian Angel):

And yet it’s been going on for 3 pages

Don’t mind me by the way, I’m only here for the entertainment

Yeah, I feel like I couldve made the conclusion earlier, but there is this morbid bit of curiosity xD 

[Lidozin]

It doesn’t really matter how many years someone has spent in DCS. What does matter is whether they took the time to find and translate the relevant section of the aircraft manual, where the correct TAS for best rate of climb is specified.

What also matters is whether they were able to fly a test profile while accurately maintaining that target TAS, and doing so smoothly, without introducing pitch oscillations or g-loading fluctuations. Even small deviations from 1g can significantly increase induced drag and reduce the rate of climb.

Now, with that done, we can simply compare the test points — both for the AI (green) and for the human-flown aircraft (orange) — directly against the official climb performance chart of the real aircraft.  Ten minutes of calm, methodical testing often yield more clarity than hours of scholastic debate.

It’s also worth noting that trying to “stay in formation” with an AI aircraft using maximum thrust is inherently flawed as a climb rate test. With both aircraft of the same type and mass, and one applying full power while the other attempts to follow, the follower will inevitably fall behind. Any such comparison will always bias against the trailing aircraft.

That’s precisely why controlled solo climb tests  flown at the documented best-climb TAS and held at near-constant 1g  are the only valid method for evaluating energy performance in this context.

[MAXsenna]

2 hours ago, Lidozin said:

With both aircraft of the same type and mass, and one applying full power while the other attempts to follow, the follower will inevitably fall behind. Any such comparison will always bias against the trailing aircraft.

Dude! You're just proving our points that the AI is superhuman, and that's why everything you write comes off as trolling and gaslighting, hence why the question how long have you been playing the game.

I guess you're gonna prove how "the all seeing eye" of the AI is also completely realistic next.  

Cheers! 

[Lidozin]

I’m not suggesting that the AI behaves perfectly in every respect — only that, in this specific context, its energy performance in sustained climb matches both the manual and computed data to within a few percent. That’s not “superhuman” — that’s simply a correct implementation of aerodynamic tables.

The formation example is a common misunderstanding:

1. A wingman falling behind during a climb is not necessarily a sign of AI "superpowers", but often a result of human-induced energy loss — especially when trying to aggressively hold position by chasing pitch and throttle changes.

2. In real flight, a lead aircraft never climbs at full power unless deliberately trying to leave the wingman behind.

If you want to stay with the AI during a clean climb, fly exactly like it does: hold a stable profile, minimize control input, and don’t chase energy with abrupt g-load changes.
This isn’t hypothetical — it’s perfectly doable in practice. In fact, in the test shown earlier, a human-flown aircraft matched the climb profile almost exactly by simply following the documented TAS and keeping the g-load near 1.0.

Finally, if there are concerns about other aspects of AI behavior — like situational awareness or detection — that’s a separate discussion, and worth having. But let’s not conflate that with correctly modeled flight performance.

[Dragon1-1]

32 minutes ago, Lidozin said:

That’s not “superhuman” — that’s simply a correct implementation of aerodynamic tables.

You're not getting it - flying aerodynamic tables exactly in an analog bird like MiG-15 is superhuman. It's like you're dogfighting a wind tunnel model, not an aircraft flown by a real human with real controls. Aerodynamics are only one part of the aircraft's performance, another critical factor is the human-machine interface. Yes, it's theoretically possible to fly the MiG-15 that way, but in practice, you'd need to build a piloting robot, or retrofit an FBW system (same thing, really, if you think about it).

A human pilot needs to physically move the control column to maneuver, actuate the trim switch to adjust the trim, physically look at the gauges to determine airspeed, keep tally, and so on. In a MiG-15, all those tasks are somewhat complicated by poor cockpit ergonomics (a somewhat notorious issue with all Soviet fighters), in addition to the normal delays and imperfections from making those actions. AI models none of this, which makes it superhuman. We're effectively fighting a MiG-15 equipped with modern FBW controls, a force sensing stick and a modern HMD.

In fact, it seems to be exactly the same case as with climbing with warbirds. Until they got WEP restricted, most people couldn't do it, but Reflected found a way by the means of unrealistically tight trimming and very hard, but doable precision flying. Yes, it follows the tables, but it does not follow either real WWII practice or normal ways to fly a warbird. This is also why complaints about fighters which are supposed to have FBW are much less frequent.

Edited Sunday at 11:39 AM by Dragon1-1

[MAXsenna]

28 minutes ago, Lidozin said:

fly exactly like it does:

Sure!  

3 minutes ago, Dragon1-1 said:

You're not getting it - flying aerodynamic tables exactly in an analog bird like MiG-15 is superhuman.

☝

4 minutes ago, Dragon1-1 said:

We're effectively fighting a MiG-15 equipped with modern FBW controls, a force sensing stick and a modern HMD.

With a pilot that turns ChatGPT green in envy.  

[Lidozin]

48 minutes ago, Dragon1-1 said:

You're not getting it - flying aerodynamic tables exactly in an analog bird like MiG-15 is superhuman. It's like you're dogfighting a wind tunnel model, not an aircraft flown by a real human with real controls. Aerodynamics are only one part of the aircraft's performance, another critical factor is the human-machine interface. Yes, it's theoretically possible to fly the MiG-15 that way, but in practice, you'd need to build a piloting robot, or retrofit an FBW system (same thing, really, if you think about it).

A human pilot needs to physically move the control column to maneuver, actuate the trim switch to adjust the trim, physically look at the gauges to determine airspeed, keep tally, and so on. In a MiG-15, all those tasks are somewhat complicated by poor cockpit ergonomics (a somewhat notorious issue with all Soviet fighters), in addition to the normal delays and imperfections from making those actions. AI models none of this, which makes it superhuman. We're effectively fighting a MiG-15 equipped with modern FBW controls, a force sensing stick and a modern HMD.

In fact, it seems to be exactly the same case as with climbing with warbirds. Until they got WEP restricted, most people couldn't do it, but Reflected found a way by the means of unrealistically tight trimming and very hard, but doable precision flying. Yes, it follows the tables, but it does not follow either real WWII practice or normal ways to fly a warbird. This is also why complaints about fighters which are supposed to have FBW are much less frequent.

What you’re describing is exactly what separates a well-trained pilot — or a skilled virtual one — from someone just “flying it by feel.”

Yes, it's hard. Yes, it takes discipline. That’s why real-world flight and combat manuals emphasize very specific energy management techniques:

Climbing at the most efficient airspeed.

Maintaining coordinated, smooth flight.

Avoiding unnecessary g-loading.

Turning at corner velocity.

Trimming properly and flying clean.

These aren't theoretical details — they're core to real-world air combat doctrine, because that’s what allows you to stay fast, stay high, and stay alive.

You don’t need to be a robot. But you do need to avoid wasting energy through unnecessary control inputs. And even if your airspeed control is only accurate to ±30–40 km/h, that’s often enough, as long as you don’t induce drag by chasing the fight with abrupt pitch changes.

As for the AI: it simply flies by the tables with clean logic and no wasted motion. That’s not superhuman — it’s what happens when someone (or something) doesn’t bleed energy.

In fact, I suspect that when some players meet another human online who does understand energy fighting, timing, and aerodynamic discipline — they’re likely to call them a cheater, too.

That said, I’d like to remind everyone that the original goal of this analysis was not to examine AI behavior in terms of tactics or input realism, but simply to test the claim that the AI “doesn’t obey physics, or has physical performance beyond what a player-controlled aircraft can achieve.” The flight test results suggest otherwise. Let’s avoid shifting the discussion away from that specific and measurable question.

Edited Sunday at 12:22 PM by Lidozin

[Exorcet]

8 minutes ago, Lidozin said:

That said, I’d like to remind everyone that the original goal of this analysis was not to examine AI behavior in terms of tactics or input realism, but simply to test the claim that the AI “doesn’t obey physics, or has physical performance beyond what a player-controlled aircraft can achieve.” The flight test results suggest otherwise. Let’s avoid shifting the discussion away from that specific and measurable question.

Indeed, however dogfights are more complicated than flying at a constant condition as in test fights. I don't think a solid conclusion can be reached from matching static testing alone. It may be that the AI doesn't pull any more g than it should while it also simultaneously doesn't feel some of the design limitations of the aircraft (as is the case for props without a doubt) or that the AI's physics violations are limited to specific actions, such as transient maneuvers or its reaction to an adversary.

[Dragon1-1]

20 minutes ago, Lidozin said:

As for the AI: it simply flies by the tables with clean logic and no wasted motion. That’s not superhuman — it’s what happens when someone (or something) doesn’t bleed energy.

Doing that perfectly, 100% of the time, is very much superhuman. We're not just talking a well trained pilot, we're talking one who can magically teleport the stick from one side to the other. We're talking one who can guess his airspeed perfectly at all times while keeping his eyes glued to the bandit (and simultaneously looking behind, the AI is that good). We're talking one who can continuously hold perfect trim, disregarding the imprecisions in the trim switch, or even the need to actuate it. In fact, I'm pretty sure nobody actually trims mid-fight. DCS AI does, it's always in perfect trim like an FBW jet.

I agree that most players don't even meet the criteria of simply flying well, and the MiG-15 is a real beast when flown well. However, no real pilot is going to hold corner speed to the knot while not looking at the airspeed dial. We simmers are seriously disadvantaged in the G load department, but a pilot's butt isn't a super-precise G indicator. You can be good at guessing, but you don't have a HMD, so in a dogfight, the MiG-15 needs to be flown by feel to some degree. Otherwise, you'd spend so much time staring at the gauges that you lose tally.

Edited Sunday at 12:41 PM by Dragon1-1

[Raven (Elysian Angel)]

On top of all that, early MiGs are also notorious for high-speed buffeting, especially in a dive - up to the point the aircraft becomes virtually uncontrollable. The AI doesn't care about any of that...

Anyway, this "discussion" / "debate" / whatever you want to call it can keep going on until the Milky Way collides with Andromeda and still won't be any closer to reaching an agreement.
There's a Mark Twain quote that comes to mind that is very applicable, but I won't repeat it here or I'll get naughty points  
Have fun, peeps! 

[Katmandu]

On 7/13/2025 at 8:11 AM, Lidozin said:

It’s also worth noting that trying to “stay in formation” with an AI aircraft using maximum thrust is inherently flawed as a climb rate test. With both aircraft of the same type and mass, and one applying full power while the other attempts to follow, the follower will inevitably fall behind. Any such comparison will always bias against the trailing aircraft.

This is a testable (as well as not at all self-evident!) statement. How about flying this test with two human controlled Mig-15s with both climbing at full power? My wager is on the difference between the leader and the trailing plane being so small as to be negligible. Certainly not the effect of one plane continuing to climb while the other is falling from the sky, like in the Mig-15 AI vs Human climb test video.

Moreover, the official document graphs were also taken form real life tests with real life pilots, not robots (i.e., a perfectly flown Mig-15 with no human oscillations would have produced better results that what we see in the official graphs).

PS A link to a well researched SFM mod of the AI Mig-15 also deserves a mention in every thread where it's mentioned  

 

Edited Monday at 08:08 AM by Katmandu

[Schmidtfire]

It's not only about AI flying the tables perfectly. The AI SFM is simplified compared to a PFM.

Perfect AI table flying + SFM = UFO result.

It all adds up, making the end result unrealistic. And this not only goes for the AI MiG-15bis, it's a global shortcoming within limitations of the AI in DCS World.

The many players that has complained about this issue for over 10 years are not wrong. Hopefully ED will resolve a lot of these shortcomings with the upcoming GFM model. And while I admire the investigative enthusiasm, I can't help but think this might be the Dunning-Kruger effect at work.

[Dragon1-1]

7 hours ago, Katmandu said:

How about flying this test with two human controlled Mig-15s with both climbing at full power? My wager is on the difference between the leader and the trailing plane being so small as to be negligible. 

Trailing plane will fall out of the sky in this case. This is because in this scenario, you're reacting to what the leader does. If you both firewall the throttles in sync, you will stay together, but what is going to happen is that you'll see the leader going full thrust, and only then apply power of your own. This will produce a lag, normally when flying formation you could correct it by adding more power and then pulling it back, but as you're both at full power, you can't. So you will fall behind/below, and if you try to maintain your position in the formation, you will start trading airspeed for altitude, putting you behind the curve and stalling the jet. If you don't, you will eventually settle into the same climb profile as the leader, but the gap that will have opened up between you noticing the acceleration and engine reaching full power (a surprisingly large delay with a slow spooling jet) will remain considerable. You would also have to only start climbing when your engine spool up, not when the leader does, as otherwise you'll climb too early and likewise end up behind the curve. Max rate climb is a very unforgiving maneuver.

Edited Monday at 03:56 PM by Dragon1-1

[Katmandu]

6 hours ago, Dragon1-1 said:

Trailing plane will fall out of the sky in this case. This is because in this scenario, you're reacting to what the leader does. If you both firewall the throttles in sync, you will stay together, but what is going to happen is that you'll see the leader going full thrust, and only then apply power of your own. This will produce a lag

But in our case the leader was ahead - as per my climb test video, the AI starts ahead. So, in his journey towards the hill, he floors the throttle first. I react and floor the throttle second - yes. But, since I was behind him, I was also further away from the hill than he - at the instant that he applied the throttle. By the time I floored the throttle I was approximately the same distance from the hill as he was. (it is also possible to be more precise, the AI starts the climb at 685km/h IAS, so it's a matter of doing the same and maintaining his angle )

Your logic would apply if we were side by side - then yes, my reaction time would delay the throttle application and I would start rolling onto the hill at slower speed that the AI "car" and thus would run out of energy sooner. There is margin for error here, but I could not outclimb the AI to 10km, no matter what profile (speed, angles) I used.

PS Anyway, I am super happy with Curly's awesome SFM mod for AI Mig-15, so it's all academic now  I only wish I found it much-much earlier...

Edited Monday at 10:13 PM by Katmandu

[Lidozin]

2 hours ago, Katmandu said:

But in our case the leader was ahead - as per my climb test video, the AI starts ahead. So, in his journey towards the hill, he floors the throttle first. I react and floor the throttle second - yes. But, since I was behind him, I was also further away from the hill than he - at the instant that he applies the throttle. By the time I floored the throttle I was approximately the same distance from the hill as he was. (it is also possible to be more precise, the AI starts the climb at 685km/h IAS, so it's a matter of doing the same and maintaining his angle )

Your logic would apply if we were side by side - then yes, my reaction time would delay the throttle application and I would start rolling onto the hill at slower speed that the AI "car" and thus would run out of energy sooner. There is some margin for error there, but I could not outclimb the AI to 10km, no matter what profile (speed, angles) I used.

PS Anyway, I am super happy with Curly's awesome SFM mod for AI Mig-15, so it's all academic now  I wish I found it earlier...

This mod is unlikely to make a meaningful difference, if only because the thrust has been adjusted in the TAS region below 300 km/h — a regime the AI almost never flies in, even at low altitude. During climb, the AI typically maintains a TAS around 700 km/h, where the thrust values in the mod remain essentially unchanged. So any claimed improvements are unlikely to impact the AI’s actual climb behavior in a measurable way.

Instead of relying on modifications, you can perform a direct test using the standard setup:

1. Place the AI-controlled MiG-15 ahead of your aircraft at a distance of 600 meters, both starting at sea level with a TAS of 700 km/h.

2. Assign the AI a route with waypoints that require a continuous climb to 11,000 meters at full power.

3. Position your own aircraft directly behind the AI (600 m), matching its speed and heading.

At mission start, apply full throttle and maintain level flight. Let your aircraft accelerate naturally until it reaches 700–705 km/h TAS, then initiate a gradual climb, maintaining 710 ± 10 km/h TAS throughout. Use trim gently to hold pitch; avoid aggressive control inputs. The goal is not to match the AI’s pitch angle, but to fly a clean, energy-efficient climb profile.

[Katmandu]

2 hours ago, Lidozin said:

This mod is unlikely to make a meaningful difference, if only because the thrust has been adjusted in the TAS region below 300 km/h — a regime the AI almost never flies in, even at low altitude.

He reworked almost all of Mig-15 SFM, way beyond a simple thrust adjustment. Thrust adjustment was my own humble attempt at a mod, this was before I found Curly's thread ( https://forum.dcs.world/topic/351306-sfm-performance-and-inaccuracies/#comment-5459098)

Here is comparison of a section of the original lua (red) and its Curly's mod (green):

[Temetre]

vor 13 Stunden schrieb Katmandu:

He reworked almost all of Mig-15 SFM, way beyond a simple thrust adjustment. Thrust adjustment was my own humble attempt at a mod, this was before I found Curly's thread ( https://forum.dcs.world/topic/351306-sfm-performance-and-inaccuracies/#comment-5459098)

Here is comparison of a section of the original lua (red) and its Curly's mod (green):

Thats interesting, so the SFM values arent actually very accurate? Implies OPs basic calculations are wrong, not just his assumptions about the application of those numbers.

Do you know if theres similar AI mods for planes like the Mig-21? Thats another classic offender.

Edited Tuesday at 11:51 AM by Temetre

[Silver_Dragon]

I still think most AI aircraft continue to use SFM, with the corresponding problems it may have, such as cases where the AI inexplicably runs out of fuel when we still have fuel left, as if they were always running the engine at 100% or using afterburner, burning fuel rapidly.

We've talked about GFM... but has anyone confirmed that any current DCS World aircraft has it implemented? We haven't heard from anyone about it for years.

As for the MiG-15Bis, I'm afraid it's an inherent problem with the old code, because it's very strange that these problems with the AI continue to exist, and that, given everything that's been said, it hasn't been minimally resolved.