Fractal Softworks Forum

Please login or register.

Login with username, password and session length
Advanced search  

News:

Starsector 0.97a is out! (02/02/24); New blog post: Simulator Enhancements (03/13/24)

Pages: [1] 2

Author Topic: Aurora as sample case for AI piloting errors.  (Read 2452 times)

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Aurora as sample case for AI piloting errors.
« on: September 05, 2019, 09:19:25 AM »

Aurora makes a great demonstrator, because it's combination of short range and high speed (but only in bursts, since most of it comes from system) magnifies effects of piloting errors.

Ai seems to mostly determine when it should approach based on relative flux (i.e. my 30% vs enemy 60% = they are vulnerable, let's rush). This is wrong in several ways:

1) If enemy has much better effective flux capacity (more raw capacity, better shield, more effective weapons), this can be suicidal. Paragon doesn't need it's full flux bar to handle Aurora, and even efficient Eagle build(see below) can easily win from higher flux percentage.
Builds like 4xTL Paragon can even afford firing when Aurora is out of range to drive up their own flux and provoke Aurora into closing in. Then kill it, once lured deep enough.

2) Tactics doesn't seem to discern hard and soft flux. So opponent can just intermittently hold fire (or use fortress shield) and let Aurora use it's inefficient weapons to reverse the initial perceived advantage.

3) Can lead to backing up after successfully disabling enemy weapons with emp, mostly when enemy has significant amount of soft flux. Opponent stops firing, their flux goes down and Aurora thinks it's time retreat (if it had decently high flux levels too). WRONG! It's time to vent in their face, since they can't punish the vent due to disabled weapons (of course need to be aware of what exactly got disabled).
(This is somewhat rare situation, but I've seen it vs same Eagle a few times).

Mostly used this Eagle vs sim Aurora with Eliminate command (otherwise fight takes too long, and will probably go into CR timeout).
Spoiler
[close]

« Last Edit: September 05, 2019, 09:24:05 AM by TaLaR »
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #1 on: September 06, 2019, 06:44:38 AM »

Exhibit B, Odyssey.
Build:
(optimized to for AI vs AI on sim Conquest, but Onslaught makes more educational opponent)
Spoiler
[close]

What I'd ideally want Odyssey to do (made a few piloting mistakes, it's possible to get behind Onslaught much faster):
Video:

What I'd be willing to accept (intentionally fighting Onslaught from front)
Video:

What AI actually does:
Video:

First, this is quite unlucky sample. Despite all piloting mistakes Odyssey usually wins, even if badly.
But all the better demonstration of what it does wrong:
1) After Odyssey disengages and dissipates most of it's flux, it always rushes forward using system, while still having about 25% hard flux. I mean cases when Onslaught has at least as much hard flux are passable, but when Onslaught has enough time to fully reset before contact? That, right there is criminal. I think at least this one should be easily fixable.
2) It doesn't even try to get behind Onslaught - obvious. While it's possible to win from front, that's much less efficient approach in 1v1.
3) Wrong choice of distance in later stages (after Annihilators run out). Odyssey loses flux war at 2xTPC vs Autopulse distance, but if it got closer into IR pulse range, it could fight under much more favorable conditions. Though at least Odyssey is willing to retreat on high flux, until Onslaught is also low armor+high flux. Which leads to next point.
4) In second half, when Onslaught gets low on armor and starts conserving flux, Odyssey's behavior becomes suicidally inefficient. Instead of decisively advancing into IR Pulse range or retreating to fully recover flux (to advance later), Odyssey just hangs within TPC attack range yet outside (or barely inside) Autopulse distance. And gets killed one-sidedly for it.


« Last Edit: September 06, 2019, 06:46:19 AM by TaLaR »
Logged

intrinsic_parity

  • Admiral
  • *****
  • Posts: 3071
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #2 on: September 06, 2019, 01:03:45 PM »

The one thing I don't agree with is the flanking stuff. AI should be optimized for fleet combat not 1v1 and trying to get behind a ship in a fleet context is almost always suicidal. I think any AI behavior attempting to get behind ships will almost certainly result in many more situations where the ship flies into the middle of the enemy fleet and dies. It's a very high risk high reward maneuver in a fleet that I don't really want the AI trying to make. Even if it seems safe at the moment, reinforcements can spawn in behind you and you're also generally isolating yourself. I don't see that behavior as valuable.

I would also point out that a big part of what you did involved turning the ship around to use the ship system to run away which is something that really only is useful/a good idea on ships with plasma burn or maybe burn drive if they can turn quickly (which most can't). I think these ships would need their own AI behavior separate from other ships for that to work (you wouldn't want your hammerhead turning around to back up), and I don't know if that is practical or reasonable. 
Logged

Alex

  • Administrator
  • Admiral
  • *****
  • Posts: 23986
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #3 on: September 06, 2019, 01:07:34 PM »

Thank you, made a note to take a look when I get a chance!

Just as a general note, a lot of this stuff is beyond what the AI can reasonably do - it has to deal with general cases; getting too specific often results in improved behavior for that one case and considerably worse behavior for things that kind of look like it but are crucially different in some what the AI doesn't pick up on. Still, definitely some things here to look at.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #4 on: September 06, 2019, 01:27:08 PM »

@ intrinsic_parity

Yes, I'd generally not try to get behind an Onslaught in large fleet fight (not with Odyssey anyway). But knowing when to try get behind would be valuable too (and once it's there Odyssey can disengage fairly safely, only approach is risky).

Squeezing maximum use out of system is very important part of piloting, and is one of areas where AI is significantly lacking. Take Hyperion - it's ALL about the system, so trying to general-case it just leads to Hyperion being unusable as AI ship. I think having some system-specific coding is unavoidable, if AI is to be good at using them.

@ Alex

Glad to be of help.

I think Aurora could definitely get less suicidal. And at least, how it seems to me, it's exactly a failure of general 'when-to-approach' logic.

As for Odyssey, of course, I don't actually expect AI to repeat 1st video. Margins are thin enough that I can get it wrong too and fail horribly. But at least charging before it finishes dissipating/venting is clear enough mistake.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #5 on: September 07, 2019, 01:23:07 AM »

Wolf vs sim Heavy Rocket Dominator, as case study for getting behind the enemy.
- Wolf can only win by doing so.
- This Dominator variant is the easiest and safest ship to get behind in whole sim.
- Even in a large fleet fight, getting isolated and killed from rear is how enemy Dominators quite often die. So it wouldn't be wise to ignore as 1v1 only tactic.

Wolf Build:
Spoiler
[close]

Player-piloted:
I went in very cautious, avoiding to shoot Reapers right before Burn drive or when all Vulcans are active.
Slow approach with shield up, without exploiting Burn Drive (still good enough):

AI steady:
It almost got CR timeout.

AI with Eliminate order:

What AI does wrong:
- Retreating too far to vent/dissipate after it already got behind the Dominator. Once you are there, Dominator literally can't turn fast enough to use frontal guns, so their range is irrelevant for considering where it is safe to vent. Correct spot in this case is right outside of rear Vulcan range.
- Repeated (and repealed) frontal approaches - these do nothing beside wasting CR. Best way to get behind Dominator is by keeping shield down for zero flux boost combined with skimmer. This Dominater build doesn't have any fast tracking guns that could inflict serious damage during such approach. But as second piloted video shows, even waltzing in with shields up works, you just need to commit to trading flux for positional advantage.
- Reaper targeting is way too risky (both land only 1 of 2), but I guess this part may be intentional.

Also, I'd say this connects to more general topic of awareness of enemy weapon arcs/ranges and hull's turn rate, it's very much possible to design a ship with vulnerable spot other than rear (Paragon often ends up exactly that without trying, sim one is easiest to attack from rear-side angle).
Logged

sotanaht

  • Captain
  • ****
  • Posts: 297
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #6 on: September 07, 2019, 01:34:37 AM »

You've made some major mistakes in your build, optimizing for the AI.  Your cardinal sin is using variable range weapons.  You have the Tachyon at 1000 base, the autopulse at 700 base, and the IR pulse lasers at 500.  Only an Agressive or Reckless AI will EVER be able to use those weapons intentionally, and considering the Tachyon is the last, not first weapon you want the AI to use, it will never use those weapons optimally.

The solution?  Ditch the IR, ditch the Tachyon, go Autopulse-only (as far as non-missile weapons go).  Plasma would be nice, but the inefficiency kills it, and it's shorter range than Autopulse so you can't use the two together.  The result will be a ship that actually hard-engages properly and fights a direct flux war, and then attempts to back off fully and vent if it loses said flux war.  In sim it might still lose to an Onslaught (the build absolutely will win with fleet-only skills, I've done it myself and the Odyssey wins with barely a scratch on it), but it's going to be a thousand times more effective in fleet settings simply because it actually gets in there and uses its weapons, and then gets out again.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #7 on: September 07, 2019, 01:43:23 AM »

@sotanaht

I'd actually take that as argument for better range band awareness for AI. Always engaging by shortest/longest range are both wrong solutions that get you killed depending on what you fight. Correct option is to engage at range-band most beneficial to you (assuming you have speed advantage to choose).

Also, this build wasn't optimized to fight the Onslaught, I just chose it as better demonstrator of AI mistakes.
« Last Edit: September 07, 2019, 01:50:06 AM by TaLaR »
Logged

Grievous69

  • Admiral
  • *****
  • Posts: 2975
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #8 on: September 07, 2019, 01:48:10 AM »

Best way to get behind Dominator is by keeping shield down for zero flux boost combined with skimmer. This Dominater build doesn't have any fast tracking guns that could inflict serious damage during such approach. But as second piloted video shows, even waltzing in with shields up works, you just need to commit to trading flux for positional advantage.
I don't think AI could ever be this ''smart'' let's say, because as others pointed it out, while it works in this situation, it'll get destroyed in fleet battles. Even if it manages to catch a ship alone, there are fighters that may come, missiles that may get launched at it. It's just too risky imo. U agree with the other points tho, it's very annoying when AI just wastes CR and does basically nothing but dance around and flicker shields.

You've made some major mistakes in your build, optimizing for the AI.  Your cardinal sin is using variable range weapons.  You have the Tachyon at 1000 base, the autopulse at 700 base, and the IR pulse lasers at 500.  Only an Agressive or Reckless AI will EVER be able to use those weapons intentionally, and considering the Tachyon is the last, not first weapon you want the AI to use, it will never use those weapons optimally.

The solution?  Ditch the IR, ditch the Tachyon, go Autopulse-only (as far as non-missile weapons go).
Solution? Lol wut the solution would be to make AI actually use weapons other than its longest range one. It's ridiculous that all of your AI ships must have weapons of exactly the same range, it makes build variety even worse.
Logged
Please don't take me too seriously.

Offensive_Name

  • Lieutenant
  • **
  • Posts: 75
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #9 on: September 07, 2019, 02:01:48 AM »

Solution? Lol wut the solution would be to make AI actually use weapons other than its longest range one. It's ridiculous that all of your AI ships must have weapons of exactly the same range, it makes build variety even worse.

I agree completely. A simple fix would be the ability to set the AI of ships without officers to Aggressive or any of the others. In the long run however there needs to be some tweaks to the officerless AI and the Steady Officer AI.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #10 on: September 07, 2019, 02:05:28 AM »

Best way to get behind Dominator is by keeping shield down for zero flux boost combined with skimmer. This Dominater build doesn't have any fast tracking guns that could inflict serious damage during such approach. But as second piloted video shows, even waltzing in with shields up works, you just need to commit to trading flux for positional advantage.
I don't think AI could ever be this ''smart'' let's say, because as others pointed it out, while it works in this situation, it'll get destroyed in fleet battles. Even if it manages to catch a ship alone, there are fighters that may come, missiles that may get launched at it. It's just too risky imo. U agree with the other points tho, it's very annoying when AI just wastes CR and does basically nothing but dance around and flicker shields.

All of these could be encompassed into decision making process (if there are fighters on field/nearby, are there any missiles in flight/can be launched, how fast can enemy turn, whether Dominator's rear is actually full packed with IPDAI Railguns to counter exactly that, etc). Though I do agree it's one of harder points to implement.
Logged

Amoebka

  • Admiral
  • *****
  • Posts: 1315
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #11 on: September 07, 2019, 02:37:51 AM »

I agree completely. A simple fix would be the ability to set the AI of ships without officers to Aggressive or any of the others.

You can already do that. Doctrine agressiveness setting affects all ships without officers, including the ones in your own battles. Put it on 3 instead of the default 2 and all your ships will behave as if they had agressive officers.

A simple fix would be to actually explain that to the player better. Right now you need to mouse over some weird bar amongst other weird bars on a tab that looks like it only affects colonies.
Logged

SafariJohn

  • Admiral
  • *****
  • Posts: 3010
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #12 on: September 07, 2019, 04:35:35 AM »

While testing for the tournament I saw too many cases of a destroyer or frigate CHOOSING to hover straight in front of a Paragon or other powerful ship when flanking was perfectly safe.

Forming a line is nice and all, but ships shouldn't be suicidal about doing it.
Logged

SonnaBanana

  • Admiral
  • *****
  • Posts: 867
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #13 on: September 07, 2019, 07:51:36 AM »


Also, I'd say this connects to more general topic of awareness of enemy weapon arcs/ranges and hull's turn rate, it's very much possible to design a ship with vulnerable spot other than rear (Paragon often ends up exactly that without trying, sim one is easiest to attack from rear-side angle).
A general preference for attacking from the direction of minimum return fire?
Logged
I'm not going to check but you should feel bad :( - Alex

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Aurora as sample case for AI piloting errors.
« Reply #14 on: September 07, 2019, 08:05:14 AM »

Also, I'd say this connects to more general topic of awareness of enemy weapon arcs/ranges and hull's turn rate, it's very much possible to design a ship with vulnerable spot other than rear (Paragon often ends up exactly that without trying, sim one is easiest to attack from rear-side angle).
A general preference for attacking from the direction of minimum return fire?

Sure, but minimum return fire is admittedly hard to define. Examples:
- for a fast frigate attacking from medium or above range, Hellbore can be greatly discounted, because it's easy to evade. While TL or Mjolnir is particularly dangerous long range insta-hit/fast projectile that's effective against both shield and armor.
- Odyssey has different priorities, all of the above will easily hit it, but can be shield tanked for quite a while. While kinetics are to be avoided. Yet, if there is a direction that is covered by multiple Needlers only and nothing else, that's also a vulnerable spot for armor tanking.
- Also, trying to reach certain facing of enemy makes sense only if you can actually do it. Against a faster/maneuverable ship, you don't get to decide. And then there are ships like Conquest, with variable turn rate - you may be able to reach it's vulnerable side (if it has flak side/etc)/rear, but only intermittently.
Logged
Pages: [1] 2