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 3 [4] 5 6

Author Topic: Do-nothing AI.  (Read 20111 times)

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Do-nothing AI.
« Reply #45 on: October 24, 2018, 09:53:40 PM »

This is a thing in 0.9, btw, or more specifically: ships that are running low on CR/peak time get progressively more aggressive.

While this is good, there is also a problem with what AI understands as aggressive play.
It's not just about reducing preferred distance and allowing slightly more flux before starting retreat.

Committed attack 1 on 1 with SO melee Lasher vs superior opponent (DE/CR) means sticking to the end, even losing 99% hull to win is better than retreating, because approach is expensive (in flux/hp) and you won't get a second chance.
Also, SO+UI also makes such Lasher fast enough to reach engines of ships like Hammerhead, and doing so is necessary to win against optimized Hammerhead variant (both skill-less). So it's about using speed/acceleration advantage to the max as well.

But all above is only good if done based on correct prediction. Charging SO Lasher at Hammerhead is doable, if hard. Trying to approach Aurora is always suicide, there is no winning tactics 1v1.

By the way, when attacking a longer ranged ship, one nice trick is to push a large asteroid in front of yourself as disposable shield. How about teaching AI to use/counter it?
« Last Edit: October 24, 2018, 09:57:49 PM by TaLaR »
Logged

Alex

  • Administrator
  • Admiral
  • *****
  • Posts: 23986
    • View Profile
Re: Do-nothing AI.
« Reply #46 on: October 24, 2018, 10:26:07 PM »

Programming the AI to do high risk/reward maneuvers doesn't seem like a good idea. That's just setting it up for failure. It'll never get a good success rate - even a good player will make mistakes making those types of decisions - and the AI won't get the benefit of the doubt the player gives themselves when they screw up. For bonus points, the consequence of screwing up here is the ship getting vaporized, so: about as severe as it gets.

It'd also make the mistakes consistently in a given failure-scenario, i.e. it'd be guaranteed to suicide the SO Lasher or whatever on a particular type of opponent it can't beat. More conservative play is generally going to be better because it'll produce less situations where it consistently makes ship-losing mistakes. Of course that will still happen but there'll be less of it and the outcomes will usually be less severe.

Likewise, trying to make the AI do something too clever (i.e. the asteroid trick) is not a good idea, for similar reasons.
Logged

Goumindong

  • Admiral
  • *****
  • Posts: 1886
    • View Profile
Re: Do-nothing AI.
« Reply #47 on: October 25, 2018, 12:19:20 AM »

Well the problem isn’t that the AI would lose ships. The expectation is that the AI is going to lose so worrything that the AIs ships blow up isn’t a great option.

The problem is that if the AI is consistently making mistakes it’s too easy to fight and so the only way to enforce a challenge is to increase numbers and the size of opponents. Then, when you do breach the cusp of what you can handle the AI will crush you fast.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Do-nothing AI.
« Reply #48 on: October 25, 2018, 12:31:30 AM »

Programming the AI to do high risk/reward maneuvers doesn't seem like a good idea. That's just setting it up for failure. It'll never get a good success rate - even a good player will make mistakes making those types of decisions - and the AI won't get the benefit of the doubt the player gives themselves when they screw up. For bonus points, the consequence of screwing up here is the ship getting vaporized, so: about as severe as it gets.

It'd also make the mistakes consistently in a given failure-scenario, i.e. it'd be guaranteed to suicide the SO Lasher or whatever on a particular type of opponent it can't beat. More conservative play is generally going to be better because it'll produce less situations where it consistently makes ship-losing mistakes. Of course that will still happen but there'll be less of it and the outcomes will usually be less severe.

Likewise, trying to make the AI do something too clever (i.e. the asteroid trick) is not a good idea, for similar reasons.

I fully agree with applying this logic to generic ships with mid-long term staying power.

But SO variants are not this. They are high risk - high reward by definition (with melee Lasher being particularly extreme example). Conservative play for them usually leads to loss by CR.
Plus, Lashers hulls are cheap, their best weapons are open market. In late-game I might want to intentionally use them as disposable pawns. But I can't - AI will just run out of CR, so I don't use them other than player-piloted.

The problem is that if the AI is consistently making mistakes it’s too easy to fight

Yeah, I think this part is unavoidable. Any sub-optimal behavior is a mistake in broad sense. Current AI already commits plenty.

Still, playing slow game with SO ships is a mistake that doesn't even need opponent to actively exploit it. SO ship just loses as time passes.
While exploiting range-management mistakes vs faster and longer ranged ship takes some effort (example: Autopulse Apogee vs Eagle/Hammerhead).
Logged

Goumindong

  • Admiral
  • *****
  • Posts: 1886
    • View Profile
Re: Do-nothing AI.
« Reply #49 on: October 25, 2018, 01:15:21 AM »

Certainly SO ships should be a lot more aggressive than they currently are. But that can probably be fixed by modifying how powerful the ships think they are rather than the AI in general.
Logged

TaLaR

  • Admiral
  • *****
  • Posts: 2794
    • View Profile
Re: Do-nothing AI.
« Reply #50 on: October 25, 2018, 01:41:58 AM »

Certainly SO ships should be a lot more aggressive than they currently are. But that can probably be fixed by modifying how powerful the ships think they are rather than the AI in general.

Precise judgement whether opponent is stronger/weaker is not an easy task, and one of things where player has huge advantage over AI.
There is definitely a lot of potential for improvement, but I don't expect AI to get it anywhere near perfect.

Still, as it seems to me, current AI doesn't even try to estimate expected outcome - it just tries to approach and see what happens (even if next instant it goes boom to 4 TLs from Paragon).
Logged

Megas

  • Admiral
  • *****
  • Posts: 12117
    • View Profile
Re: Do-nothing AI.
« Reply #51 on: October 25, 2018, 06:26:28 AM »

There is one other use of Safety Override:  on haulers for maximum speed to flee in a pursuit, where reduced CR is irrelevant because they will escape quickly if not stopped (and they will not stop or break off to fight).  Aside from SO Lasher, I use Safety Override on all of my small freighters that can take it (i.e., Hound and Cerberus).  Of course, such ships are so OP-starved that I do not put any weapons on it, even PD.  (Although if they had lots of OP, I might arm Cerebus for PD.)

If not for that use case, I would not mind having normal personality overriden to be Reckless for Safety Override.
« Last Edit: October 25, 2018, 06:35:17 AM by Megas »
Logged

Schwartz

  • Admiral
  • *****
  • Posts: 1452
    • View Profile
Re: Do-nothing AI.
« Reply #52 on: October 25, 2018, 07:24:35 AM »

Or giving ships with no officers the option to set behaviour ranging from timid to reckless. The upside of an officer is all the skills, downside is that they decide how brave they're gonna be.
Logged

intrinsic_parity

  • Admiral
  • *****
  • Posts: 3071
    • View Profile
Re: Do-nothing AI.
« Reply #53 on: October 25, 2018, 12:04:51 PM »

Honestly, you probably shouldn't give the AI SO. It's just going to either result in highly suboptimal play, or semi-frequent suicide. I'm not sure it's really possible to avoid the issue for ships that require such high risk commitment to get any value. Even the player has trouble accurately assessing that.

Well the problem isn’t that the AI would lose ships. The expectation is that the AI is going to lose so worrything that the AIs ships blow up isn’t a great option.

We are also talking about friendly ships under AI control. I'm not ok with my fleet dying because the AI makes silly mistakes. If I lose ships, it should be because I made a tactical or strategic error, not because of randomness associated with the AI.

I think a big issue is that you're fighting both with and against the AI, so any change that improves the behavior of your own fleet also makes the enemy more difficult. You want your friendly ships to play safe so you don't lose them, but then it may be frustrating to play against the enemy when they are playing very safe.
Logged

Thaago

  • Global Moderator
  • Admiral
  • *****
  • Posts: 7174
  • Harpoon Affectionado
    • View Profile
Re: Do-nothing AI.
« Reply #54 on: October 25, 2018, 12:38:29 PM »

Rather than an individual AI change, this could also be approached from the "commander" level AI. The enemy could start using a lot more eliminate commands on vulnerable player ships. Basically any time the AI has 2:1 or greater (of equal weight ships) in a local situation it should be slapping down an eliminate order.

Letting the player know about this in some way (something that for example the Combat Chatter mod could handle) and letting the player then counter by bringing in reinforcements and picking off the AI ships that are now reckless sounds like good interactive gameplay to me!

The AI commander should eventually remove the eliminate order when the target ship is protected as otherwise will lead to silliness (all their ships charging into the reformed player deathball) but allowing for some time to counter would imo be good.
Logged

Megas

  • Admiral
  • *****
  • Posts: 12117
    • View Profile
Re: Do-nothing AI.
« Reply #55 on: October 25, 2018, 01:09:13 PM »

If enemy AI has limited CP like player, they may not have CP to spare.  They burn their initial CP to capture points, and they may need to save some to retreat their ships that you will inevitably smash.
Logged

Goumindong

  • Admiral
  • *****
  • Posts: 1886
    • View Profile
Re: Do-nothing AI.
« Reply #56 on: October 25, 2018, 01:15:10 PM »

We are also talking about friendly ships under AI control. I'm not ok with my fleet dying because the AI makes silly mistakes. If I lose ships, it should be because I made a tactical or strategic error, not because of randomness associated with the AI.

Similarly to how we aren’t concerened when the enemy losss ships. Friendly AI and enemy AI do not have to be the same
Logged

intrinsic_parity

  • Admiral
  • *****
  • Posts: 3071
    • View Profile
Re: Do-nothing AI.
« Reply #57 on: October 25, 2018, 04:01:33 PM »

I don't like if the enemy does stupid things either even if it benefits me. The AI should behave logically, but imperfectly so that the player has a chance. The enemy behaving illogically is unsatisfying. Also I think the core of the AI is mostly the same for all ships, but I could be wrong.
Logged

Xaiier

  • Lieutenant
  • **
  • Posts: 56
    • View Profile
Re: Do-nothing AI.
« Reply #58 on: October 25, 2018, 04:03:10 PM »

Honestly, I would be fine if friendly ships died as a result of aggressive actions that actually meant something. We're fighting battles here, it's not realistic to expect every engagement to result in a full sweep with 0 casualties. What's frustrating is when the AI dies because they are acting stupid and treating my orders as "suggestions". When I say "eliminate", they should do so. When I say "retreat" or "avoid", they better listen, and not dawdle around. 99% of the time when ships die it's because they are just acting dumb, and this kind of "durr" behavior usually cascades into total annihilation because at a critical moment one of them wimped out instead of taking one for the team. It's very frustrating to see something like this happen and watch as your entire fleet is slowly eviscerated, because there's nothing you can do to stop it, because they won't listen.

tl;dr the default behavior is ok-ish, but I would like to be able to influence their behavior in a stronger way. That way I'm the one who makes the decision that can get a ship killed, which keeps the AI from being responsible for that kind of thing, and also allows me to exploit situations I see that the AI cannot feasibly be smart enough to take advantage of.
« Last Edit: October 25, 2018, 04:08:45 PM by Xaiier »
Logged

ThePollie

  • Lieutenant
  • **
  • Posts: 64
    • View Profile
Re: Do-nothing AI.
« Reply #59 on: October 26, 2018, 09:24:28 AM »

If you have an overwhelming advantage how do you lose?  When the AI has an overwhelming advantage it will tend to push it

Because I have six destroyers and twelve frigates trying to jump an overextending cruiser. But rather than push that advantage, I get mobbed by the entire enemy fleet before noticing my fleet decided to quickly 180 and fly to the furthest corner. Because *** gameplay, I guess.
Logged
Pages: 1 2 3 [4] 5 6