Because if they were allowed to diffuse into other systems, the majority of the PVE player base would likely cry the game is too hard. Yes, this request does not make the game realistic, but neither do watered down NPCs. This request, if honored, would create a tiny spec of realism, that is all.
Request for a SINGLE Solar System with sophisticated AI NPC flight model
- Thread starter zeta
- Start date
That is an interesting point.
Training the AI in a short amount of time may take a lot of resources. Once it is trained, to run it may not.
For example, AlphaZero (https://en.wikipedia.org/wiki/AlphaZero) is neural AI developed by Google that defeated stockfish, one of the best 'brute force' chess engines. It was not a fair game, because stockfish was not allowed to run on serious hardware for the match, while alphazero was. The reason they did this was to even the field and demonstrate that alphazero can learn chess very quickly, in 4 hours, given sufficient hardware.
Then Lc0 (Leela Chess Zero http://lczero.org/), the open source implementation of alphazero (alphazero moved onto protein folding problems) is now on par with stockfish, Lc0 runs on GPU.
The point is, it may be possible to train multiple levels and styles of AI NPCs according to the particular build of the ship, and even if training may be resource intensive, running it may or may not be.
Last edited:
I agree that the ineptitude of NPCs is an issue but at least it is universal.
The single shining light of competent NPCs to me loses as much realism because it is a single system, the other issue with this is if this was introduced and was also popular the perennial problems encountered by many of us in high traffic systems of low framerates, lag etc. would raise their ugly heads and ruin any fun to be had.
Maybe. But i think that any static upgrade of the AI will in reasonable time result in what we have. We might have a hard time defeating the AI for a limited time. But as long as it does not change and adapt, we will figure out what it does and adjust our own way of fighting accordingly. And the, in no time, start complaining again, on how boring the game is and that the AI is simple and uninteresting.
Sidenote: At the same time, while the really dedicated part of the community will adapt and overcome the new challenge and find it trivial in some time, a bit part of the community, which plays merely for fun and does not want to get the highest challenge will be left in the dust. Considering that playing for fun is a very reasonable way of looking at a game (and move away the moment it is not fun any more makes absolute sense), and that this kind of players is most likely the majority of the playerbase, the experiment would not very likely do more damage than good, too.
However, that is subjective - what you or I find easy or not necessarily challenging (especially if fully engineered), others may find the opposite. My father, for example, is in his 70s and just can't cope any more with higher difficulties - I'm hoping to be able to get him in the game with the introduction of Odyssey as he can cope with FPS gameplay far better than he can ship flight/controls/interdiction/combat etc....but there's no way he could cope with even the current NPC threat levels (with respect to current ship-based gameplay) without some difficulty and much frustration. While I would like to see threat levels to be a bit higher in anarchy systems (they should be like the lawless wild west after all), it is not universally true that everyone finds the current level of NPC threats easy or as you put it, 'a joke'.
Doesn't need to be in the 70s. Just like a number of gamer friends i still have from university time, we're on our 40s. And when i got them to try ED, several of them were freaked out by the games complexity. They simply but very understandably told me that between work and family, they can't invest that much time into learning to play such a complex game. So we stick to "simple and fun" instead.
And that happened already years ago, when NPCs were still much softer than they are now. Seemingly unknown to many here, our NPCs were ramped up at some time by giving them more engineered components. Which makes entering the game even harder.
I mean, i do get it: for us veterans in fully engineered ships, the NPCs are mostly a joke. I can solo wing assassination missions and do so for the mere sake that they, unlike almost any other content, at least are challenging enough for my Krait MK II that they still can force me to retreat sometimes and almost always leave me with at least some hull damage. On the other hand, the last times we got new ships (with NPCs already buffed up), i also merely took the new ships out to fight, without doing any engineering. I went to HAZRes and picked targets of all size and combat rank, and never ran into any issues.
So it might be that even i learned something in the run of the years. Compared to many pilots who fly for years, i perform badly. (No discussions about that, i sometimes watch some youtube videos, where i find that i would need weeks of dedicated training to merely reach that league of piloting. Which my schedule of once or twice a week, late at night, when my wife is in bed, just does not allow me to get. ) At the same time i am a vastly superior pilot to new players. As mentioned somewhere else, some of my friends actually around X-mas returned to the game, and compared to many of them, i suddenly seem to be a god of piloting.
Which kind of puts things into perspective: i am both vastly superior in skill to my newly returned friends but also far below the skill level of many other pilots around. The game allows a massive scope of piloting skill, and unfortunately it seems like many people posting here seem to not even realize how far above and beyond the capabilities of the average player they are by now.
Last edited:
To be clear, I'm not advocating for killer NPCs across all systems just where it thematically makes sense.
I sympathise with the difficulty curve to get into the game, but the difficulty should increase without the player deliberately ham-stringing themselves.
Thargoids are, I believe, the ultimate endgame NPC - does your father attack and destroy them?
I take these considerations to heart. I'd like to reiterate that my suggestion asks for ONE system out of many billions to provide a challenge. All PVE/PVP players should be unaffected by this change. That includes the elderly, the family men, and anyone else who does not have the time to overcome the steep learning curve. In fact, I would not put forward or support any change that reduces the player base, we are here to have fun together after all.
On another note, I am probably one of the worst pvpers you meet.
Yet I can take an unengineered fdl into a CZ and kill the spec alpha ship relatively quickly (4 PAs, 1 cannon, shoot, FAOFF boost flip, boost towards target, rinse and repeat).
In time I believe most of us could do that (I am 45, I use two sticks, I am not a gamer, not dextrous, but spent 2 years learning FAOFF, probably 2-3 hours / day).
I may be incorrect but I believe machine learning is sophisticated enough to accommodate different flight skill levels of NPCs.
I also believe they could be adaptive. It would be nice to hear this confirmed, or disconfirmed, from the DEVs.
You basically just wrote there that you spent between 1460 and 2136 hours of practicing FAOFF, so just one of many aspects of piloting a ship in ED, but still consider yourself to be nothing special and everybody could easily get to that level by practicing a little bit. You might want to re-evaluate where you seem to actually stand in terms of piloting skill.
On the machine learning: we spoke about that above. In the end, it's only the Devs who could give the final answer, but i'd yet have to see any game where machine learning for th NPCs worked out well. (Though, i am not aware of any examples where it was really tried on a professional level. ) I mean, no matter how sophisticated it would be, if it was "fixed" in code, somebody might merely invest 1460 hours to learn how to beat them, then find them trivial and nothing special. (Sorry, but you yourself make a very good example here. )
If it would learn and adjust, it can probably keep up the challenge. Depending on how it's done, if it would learn with the player, scaling might not ever be a problem. Hard to tell, before seeing it in use somewhere. And hard to imagine how it could be implemented well, without singifically ramping up infrastructure costs. (Something which FD tries to keep as low as somehow possible, while even limited investments early during the games development could have improved several aspects of the game a lot. )
And the last point, on the "one system only", that would merely create a hotspot, contrary to the whole games design up to now. I don't think it would make good sense. The often mentioned scaling according to security levels (so in anarchy systems the pirates would use the highest level of AI, while in high security systems the police would have it) would make sense.
There is some merit and potential to this idea. Mind you, the AI also already now has different levels, easily recongizeable on the combat rank, but unfortunately up to now all places seem to have the same distribution of NPCs of all ranks.
Last edited:
OP, i think your suggestion is reasonable. Maybe a mini bubble 10 system bubble 500 ly out where the NPCs are designed to emulate players. Call it the Legion Region or something catchy but instantly understandable. Make it so people seldom go there by accident and get caught off gaurd. I think the programming hurdle is somewhat of a cost concern though. Making NPCs aim good seems to be easy. Making them learn and be aware of their ships comparative strengths and weaknesses sounds much more involved. Overall i like the idea though.
Right, when you put it that way, I spent 1460-2136 hours, or 5256000 to 7689600 seconds on FAOFF, a heck of a lot of secondsYou basically just wrote there that you spent between 1460 and 2136 hours of practicing FAOFF, so just one of many aspects of piloting a ship in ED, but still consider yourself to be nothing special and everybody could easily get to that level by practicing a little bit. You might want to re-evaluate where you seem to actually stand in terms of piloting skill.
On the machine learning: we spoke about that above. In the end, it's only the Devs who could give the final answer, but i'd yet have to see any game where machine learning for th NPCs worked out well. (Though, i am not aware of any examples where it was really tried on a professional level. ) I mean, no matter how sophisticated it would be, if it was "fixed" in code, somebody might merely invest 1460 hours to learn how to beat them, then find them trivial and nothing special. (Sorry, but you yourself make a very good example here. )
If it would learn and adjust, it can probably keep up the challenge. Depending on how it's done, if it would learn with the player, scaling might not ever be a problem. Hard to tell, before seeing it in use somewhere. And hard to imagine how it could be implemented well, without singifically ramping up infrastructure costs. (Something which FD tries to keep as low as somehow possible, while even limited investments early during the games development could have improved several aspects of the game a lot. )
And the last point, on the "one system only", that would merely create a hotspot, contrary to the whole games design up to now. I don't think it would make good sense. The often mentioned scaling according to security levels (so in anarchy systems the pirates would use the highest level of AI, while in high security systems the police would have it) would make sense.
There is some merit and potential to this idea. Mind you, the AI also already now has different levels, easily recongizeable on the combat rank, but unfortunately up to now all places seem to have the same distribution of NPCs of all ranks.
!In all seriousness, I meant it as an encouragement for anyone who aspires to obtaining that skill. Trust me, you cannot be much worse at gaming than I am, and yes it is a long road, a time sink, one probably could spend their time better, but there is light at the end of the tunnel and for me I feel like it is good for my brain ! You can do it with a series of assassination missions which make it fun, if you are interested how let me know.
Regarding varying level scalable AI NPC flight:
If the NPC learns from the best PVPers, even though it may be non-adaptive, it may beat almost all of us for quite a long time, and perhaps it can be watered down.
Based on my observations when I do assassination missions, the pirate lord FDL flies identical to the deserter python, and these two are very different ships. Even if they hard code the flight model, they should at least take into account whether the ship has an oversized distributor, how agile it is, ... . This is not a matter of how difficult / easy the NPC flight models are. I am not certain there are flight modelS, there may only be a single flight model for all medium ships and all loadouts. I think for a spacesim that prides itself on its flight model, we can do better and I hope all, including PVE/PVP players would benefit from this versatility ?
It sounds interesting and awesome. But based on my general flight skill, how much time i can invest (and as said, usually late at night, when i am not in mint shape any more) compared to how much time you invested, it won't ever make sense to me. I can handle FAOff for tight turns and some maneuvers, might even pull one of two shots off before i have to stabilize a bit again by going FAOn and that's actually good enough for me. Merely being able to do so, without considering pip management and a number of things which can be done in FAOn i experience to be above anybody of my casually playing friends.
I know you mean it positively and want to encourage people. But it's not that easy. I see myself as dedicated player. ( Why else would i still be around since launch, manage to persuade some friends again and again every other year to give it a try again and spend some time in the game almost every week? ) Yet i don't see myself able to invest as much time as you describe that you've put into learning just one aspect of space combat, and my time budget for gaming is still above and beyond many of my friends, who also can't do my usual late night sessions, as they know that their kids will be up again and requiring attention early in the morning.
So take it as it is: you believe to not be a great pilot, but everything you wrote tells me that you actually, by investing time to train, are well above the level of most players, who just play to have fun. ( Which is quite a resonable way to play a game, no way to fault them. We more dedicated people should be happy they are around. Casual players and their money usually is what keeps games alive. )
Hmm. First, "the best PvPers", so you'd limit it to learn from a selected few. But while those people might be really good, any limited training session, especially with only a few people, will almost certainly result in flaws in the AI. (There are plenty of examples around on AI project gone wrong. Im many cases, the root problem was flawed, incomplete or flawedly filtered input data. )
Next, flight model is different for all ships. And the AI is limited by different flight models. It's still possible to spend a time in a big ships blind area now, as NPCs don't reverse as much as players. It's much harder to do so against more agile medium ships. So the ships flight model definitely matters.
What you actually refer to is not the ships flight model, but the behaviour pattern of NPCs. In this case you are basically right. There is one AI behind it, which merely learns additional things at higher combat rank.
A number of these things can be figured out when using a SLF. Put an NPC in your SLF, have it duel with the SLF of a friend. Then during the fight you switch ship with the NPC. The first thing you will notice is that NPCs with low combat rank always uses a rather bad pip settings and will not change them. While the high ranked NPC will adjust depending on what just happened.
Of course, they never go like "the enemy will probably in a second or two have me in his firing arc, so i now put pips to shields, like a good human pilot does. Experience and prediction of events to happen is what we humans can do much better. But switching 4 pips to shields in a fraction of a second when taking damage is what the computer can do quite well. Depending on the combat rank, our AI can do that, but the medium-ranked AI seems to need several secods before switching pips around, while top rated AI seems to do it within a second or so.
That all being said, the AI seem to basically use the same "set" of movements (again scaling up with the NPCs combat rank), no matter which ship it is in. The ships flight model then takes over, limits what it can do by traing to do those movements and how it actually turns out. So yes, basically all ships of the same combat rank always try to do the same set things. Which is why it gets quite predictable for us. But to be fair, having a different AI for each ship would be a nightmare. Not so much to code, but to maintain. It would mean that every ever so slight change to our ships would mean that the coder has to adjust things at over 30 places. Which also means over 30 occasionals to make a mistake, miss a detail, twist some numbers or have some other typo.(And that's before considering the test process afterwards, where any change of the AI behaviour might require some hours of testing to be sure that nothing goes wrong. Now multiply that by the over 30 ships we have now... not nice. )
To come to a conclusion: i think that within the scope of our ships and the limits of handcrafted NPC behaviour, we're at a rather good spot. I see what the NPCs were able to do and sometimes still can pull off on my friends with really low playing time in the game. I can see several aspects where combat rank makes their behaviour better (see: try with your SLF and NPCs of different ranks) and while i can't confirm it by that method, think that there are some more aspects where the combat rank has some good impact.
So yes, of course it could still be improved. It basically always is possible to improve software, add something new and fancy, etc. But i am not sure that it would improve the game too much for the average player. (I know you still consider yourself to be not that great. But merely by what you can teach to me in FAOff, even if you'd not have learned anything else while playing the game, you'd be far above the average. ) We long time players here are oddballs. A stronger AI might be interesting for us, but i don't think that it would be a big improvement for the game at large. (In case of doubt, look at some of the threads where nerfs were asked for. Where players who are around since years fight tooth and nail against any nerfs, especially to their ships defense. Imagine how somebody who already threatens to leave the game if his shields resists were nerfed by just a few percent would react to an actually cunning AI... )
And that's still on hardcoded AI. On machine learning, it would indeed be an interesting experiment. But as said, a limited data set of people to learn on, just as well as a static set of behaviour coming out of a time limited learning session would probably merely leave us with NPCs behaving differently in combat, but we players would still learn how to handle it. And soon enough we veterans would know it all again and ask for improvements again.
(While the average player is left in the dust forever. )
Last edited:
I don't think it would be necessary to give each ship its own unique profile, and in the long run it wouldn't even provide that much extra variety for players as they simply get used to how to fight each ship and different ships are already somewhat different to fight even on the same AI model. More difficult, yes, but ultimately it'll just be "oh, another Cutter, expect reverski and jousting for the next couple of minutes". A per-ship AI system also wouldn't account for differences in potential loadouts, as some ships can be outfitted for different fighting styles quite effectively, plus it wouldn't account for the player's ship and loadout as smart manoeuvring also depends on your opponent and not just your own loadout. Per-ship AI would also complicate things as more ships are added, as it would mean that every single ship FD adds into the game would also require special AI tweaks in order to work properly, rather than FD simply being able to plug an existing script into the new ship.
Rather than having a fixed per-ship AI system, the AI could instead look into relevant ship variables, potentially comparing them to their target for higher ranks, then running through an algorithm in order to figure out an effective strategy. If an AI ship has high manoeuvrability compared to the player, they would tend to opt for trying to outmaneuver the player and stay in blind spots, while a lower manoeuverability ship is more likely to rely on jousting or even reverski to keep a player ship in their sights. A shield based build would prefer to have periods of extreme range to their target to give them a chance to regenerate shields, particularly if they have biweaves, while a more hull-focused build or one with prismatics would be more likely to remain in short ranges in a damage race as they can't win in long-term attrition. A ship with a hangar bay would be happier to disengage than one without as it can still contribute even from several km away. Just a basic algorithm comparing relative strengths and weaknesses and getting sufficiently skilled NPCs to try to play to their strengths would go a long way
This would obviously involve the NPC's combat rank to help determine their strategy, but could also be expanded to include a series of per-NPC personality variables or by sub-dividing the combat rank into different categories to give even same-ranked NPCs their own fighting style. For example, a NPC with good marksmanship but poor helmsman skills would likely opt for reverski or "floating turret" styles of play, even if their ship isn't particularly suited for that style of combat. As another example, a highly reckless NPC might be quite happy to attempt rams on a player or spend prolonged periods at shorter ranges, while a cautious NPC would prefer to stay at longer ranges and be happy to disengage to regen shields even when they are still going strong.
In terms of neural network AI, I quite agree that training it on a select few players would cause problems by picking up bad habits or not knowing how to counter basic strategies (Starcraft II neural network AIs trained on grandmasters tended to perform poorly against weaker players as it didn't understand them or their strategies), as well as causing it to potentially rely heavily on abusing or countering broken mechanics rather than actually being good in its own right. Instead of being trained against players for use in the game, I see the greatest strength of a neural network AI to actually be used in balance testing. It is obviously quite challenging to figure out imbalances from in-game data as they are strongly affected by meta mechanics, while an AI would have no such issue and would instead be entirely based around in-game mechanics. By seeding a pool of independently growing and developing AIs and pitting them against each other, potentially separating them into independent sub-pools to prevent a single global meta from forming, FD would be able to identify any overpowered combat mechanics. For example, if every single sub-pool evolves the same metagame, then that's clearly a balance issue as each AI "focus group" has arrived at the same conclusion on what the strongest strategy is; however, if each sub-pool evolves a radically different meta, then it's a sign of a balanced game as no single strategy dominates, even more so if each sub-pool of AI maintains a whole menagerie of different strategies.
That part indeed sounds very interesting. Unfortunately i also fear that the effort would outweight the gain, especially when considering how much focus FD has put on balancing up to now.