Hide From the All-Seeing AI

---

Introduction

► Timestamp 00:33
The status of the AI in DCS, generally speaking, is poor. It does not use any sort of tactics, neither offensively nor defensively. All it does is notching or turning cold when a missile enters their “awareness bubble”, which is around 10nm. The flight model used by the AI is simplified and often inaccurate, looking at you, MiG-21. ED has introduced marginal improvements, but we are far from a believable representation.
Although what I just stated is common knowledge, the situation is not as bleak as it sounds if certain measures are applied. The AI, in fact, can become slightly more realistic if mission designers and players understand what makes the AI’s situational awareness borderline unlimited.
The scenario used in this discussion is elementary. A first AI, called “drone” going forward, is spawned and switches radar and other devices off. After 15 seconds, another AI tasked to execute CAP is introduced. The objective is to evaluate when the latter aeroplane engages.

Player’s actions: Radar & ECM

Let’s start by checking what players can do to make themselves less visible.

Radar

► Timestamp 01:55
The radar is frequently compared to a torch that allows its user to see in the darkness, but everyone else can see it too. The reality is possibly worse: as we have seen in the Electronic Warfare introduction I posted back in 2022, the radio waves not only have to cover the distance from the radar to the target, but also come back. That’s why, among other more complex reasons, a radar can be detected well beyond the range at which the radar itself can spot the presence of a potential target.
In this scenario, the drone activates its radar after 30s. The reaction of the Ace AI is instantaneous.
So, intuitively, players should switch the radar off when not necessary. For example, when I am playing the F-4E, the APQ-120 is always on standby if the target is beyond 50nm. There are, of course, several caveats, such as radar characteristics and type, mission and tasking, presence of controllers, et cetera.

Electronic Countermeasures

► Timestamp 02:52
ECM devices in DCS are all noise jammers, covering 360° and reaching significant ranges. Unfortunately, EW, acronym for Electronic Warfare, is represented at a very basic level in DCS.
This test shows whether activating the ECM can help the AI steer towards the drone. I placed the usual CAPping Tomcat, and I measured when a drone was intercepted, which is at circa 94.3nm. Next round, I had the drone activate the jammer right away. The results are a bit baffling, to be honest.
So, noise jammers in DCS hinder the range determination. Angles are alright, though, so a target can be tracked. The thing is, any RIO can eyeball the target’s heading using drift. The AI instead decided to… turn around and go home. I suppose that the logic here is that since the range cannot be determined, it’s not worth committing. However, an aircraft with its jammer activated is discernible from even further, and a section can triangulate the angles.
As confusing as it sounds, it appears that using the ECM pre-burnthrough might even be beneficial to “hide”, so to speak, from the AI. I’m quite sceptical about this, though.

AWACS

► Timestamp 04:13

Terrain Masking and Altitude

► Timestamp 04:40
Terrain masking, as the name suggests, involves the usage of terrain to approach a target without being seen. This method works even against the AI, but with some caveats, since the modelling of the field of view is less refined than, for example, Heatblur’s Jester.
In fact, we can observe absolutely silly scenes where the Attacker AI locates the target via radar in a nanosecond, which makes no sense since the radar should have a certain setting in terms of bars and azimuth. On the other end, it seems that after circa 60 ATA, the AI stops seeing the drone. More importantly, line of sight is required and, if the drone manages to hide again, the attacker drops the chase. Which is silly in another way, but hey, given the issues with the radar, I’ll take it.

Unfortunately, here comes another nonsense: pulse radars have full look-down capability when used by the AI. The following is an example just to demonstrate this point, and it is flown by a Rookie F-4E. It reacts in the same way as the Tomcat.

► Timestamp 06:09
So, just for a laugh, I decided to try the scenario myself from the backseat of another Phantom. Apparently, I am a terrible WSO because I am incapable of locating a target flying 24,000ft lower through the mountains, at 40R and with an antenna elevation inversely skyrocketing from -15 to -22 in a matter of seconds. Damn folks, I’m worse than a Rookie AI it seems!
…this is so sad…

Contrails

► Timestamp 06:43
Contrails, or chemtrails if you believe Earth is flat, are usually created by the engines at a certain altitude and in certain atmospheric conditions. In DCS, they are visible from considerable ranges. The AI, however, does not seem to notice them, and engages targets only within circa half a dozen nautical miles if no radars are involved.
Hopefully, they will be taken into account when the new DCS weather is completed.

Radar Notching

► Timestamp 07:14
Without opening the Pandora’s vase of what notching is and how it is represented, I demonstrated a few years ago how a player can disappear from the AI’s radar by notching them. This is an image from my manual that shows the effects. As you can see, it’s quite neat.
Since there is a lot still to say, I won’t replicate the example, but it’s rather simple if you take advantage of an external agency to monitor the Antenna Train Angle.

AI’s Field of View

As discussed, the AI does not use a radar, but rather an apparently range-based bubble to look and engage targets. However, such a bubble is not a complete, immutable, sphere: as we have seen, it interacts with the terrain, and this scenario shows how the AI cannot spot targets flying behind it.

IR Engagements

► Timestamp 08:05
Interestingly enough, the drone can even engage the AI with IR missiles and never get caught. This is excellent, as it means that the AI can be really surprised, assuming players do not broadcast their presence to the whole world by having their radar constantly emitting.

Activating Radar

The next scenario shows, in fact, how quickly the AI reacts to the drone activating its radar. As usual, the problem of the reaction time presents itself again.

AI Formations

► Timestamp 08:53
This example shows how the AI somehow respects the blind spots. However, it does not appear to transmit information if the two aeroplanes are in separate groups.

External Factors: AWACS, GCI, wingmen

► Timestamp 09:24
In an old video where I discussed the bugs and issues of DCS’ AWACS, I mentioned the impact of real radars, especially older ones, compared to the precision of what we have in the game. We have seen what happens when an AWACS is added already, but let’s review and expand the example by comparing it to ground-based intercept controllers.
In the meantime, it is interesting to note that there are probably more AWACS in any DCS server than in real life. The number of E-3s should be, for example, less than 70. Ok, let’s do 69… Anyway, not only are these assets extremely valuable and protected, but they are also nowhere near as common, and they are an almost exclusive asset of NATO and partner nations. So, a good first step is to curb their numbers in the missions, especially as the scenario is set further and further back. Besides, as occurred during the Kosovo conflict, AWACS in real life do not see everything until a de facto fixed range, and other assets may find and escalate before they can assess the situation.

Back to the testing scenario, the AI immediately reacts to the information provided by the AWACS, with superb reaction times. The AI reaction times are, in fact, another big issue that I raised in a dedicated video.
So, one solution is to reduce the number of AWACS and introduce what, for the longest time, has been the primary guidance for aircraft around the world: Ground Control. Unfortunately, here we run into the same problem again: lack of realism.
Let’s consider the following example. This passage (timestamp 01:11) is taken from the book “The MiG Diaries: Fighter pilot memoirs & accounts of Cuban, SAAF and Angolan air combat in Southern African skies”. The authors are Lionel Reid and Eduardo González.

The situation described is vastly different from what we have in DCS, and I would love to see the uncertainty factor affect both controllers’ radars and AI. Such a feature would go a long way towards increasing realism on every level. Nevertheless, Ground radars are static and immobile, and impacted by terrain masking, thus reducing the symptoms. However, this is almost a palliative, not the best solution.

SAM Radars

► Timestamp 13:41
Before moving to the next chapter, I tested whether the radar used by SAM batteries could trigger AI fighters. I verified both the SA-2 and the SA-10, and neither of them worked. Spawning a standard EW radar, instead, did the trick.
This is good news, as players can tailor scenarios without having AI fighters go “eye of Sauron” mode when a couple of air defences are around.

Biggest issues & Suggestions to ED

► Timestamp 14:07
As we have seen in the examples above, the AI has extremely short reaction times. If the player’s radar is emitting, the AI immediately reacts and also seems to know the separation, which is quite absurd. Modern RWR can assess the distance, but an SPO-10 really shouldn’t. Still, the reaction is the same.
RWRs have different capabilities. At least a two or three-tier system should be implemented, approximating the ability of the RWR to assess range, direction, and other parameters. Seeing a MiG-21 sporting an SPO-10 nailing the target’s location instantly is quite disheartening. Such an approach to technology levels should be applied across the board. At the end of the day, a perfect representation of hardware discrepancies may not be possible due to how resource-intensive it would be when applied to the AI. However, I am sure every player would love to see minor improvements to make the AI more different and unique.
Next, the radar coverage is a bubble based on range and a blind spot in the target’s rear. We desperately need a better system to simulate the AI’s radar. Let’s take the F-14’s AWG-9. Ignoring the realism of having it set to 8 bars and ±65°, the whole scan takes circa a dozen seconds. Moreover, 8 bars are not sufficient to cover the entirety of the vertical space within a few miles. A quick fix would be the addition of delays on top of a skill-level-based fixed timer. After the timer, a second check would confirm the return. Something akin to a contact “debouncing”, if you will. Something as simple as this would give targets appearing for an instant through the terrain a chance to hide again, for example.
Again, such a timer could be adjusted on a skill-level basis, something that does not occur at the moment.
Before moving to the conclusions, it is worth reiterating the issues primarily affecting the AWACS. In one of the videos mentioned earlier, Bignewey mentioned that they would have looked into the issue, but I am not sure any appropriate action has been taken yet.

Conclusions

► Timestamp 16:39
So, what to think about DCS’s AI? In primis, it is not as absurd as most people depict it. Their radar performance is ridiculous, but the visual lookout is acceptable. The problem is the addition of many parameters that, eventually, make the AI omniscient. To be fair, it’s not that players do not enjoy an excessive and gratuitous level of situational awareness, especially when we move from the Cold War era towards modern computer-flown aeroplanes – see the Hornet’s Radar Warning Receiver. What is worse is that different modules follow different quality standards, but this is a point I keep repeating since I created FlyAndWire.
At the end of the day, those issues primarily affect folks who enjoy single-player or PvE content. The thing is, SP enjoyers are the vast majority of the players. Ergo, these issues, both related to the AI and mission making, probably deserve more attention.

---

Leave comments and feedback in the Discussion Forum thread.