Jump to content
Sign in to follow this  
Drakon

Advanced AI - Public Question to the Goldhawk Team

Recommended Posts

I asked in the Xenonauts 2 Kickstarter Stretch Goals thread at about which sums we can expect better AI that will be able to prove a challenge due to strategic / tactical decisions rather than merely stat superiourity and larger numbers - a question that was echoed by one of the next posters, yet there was no answer. You chose to hide all replies to that thread and lock it, so i ask again here.

Edited by Drakon

Share this post


Link to post
Share on other sites

I think you're probably already aware that the question you're asking can never have a definite answer, because even if it was possible to accurately cost out development it's not possible for me to know exactly what standard of AI you would consider a challenge. It's obviously different for everyone; it's like asking me how much money we would need to have "good graphics" ... really, it depends on what you consider good graphics to be.

The reason I mention this is because I remember an exchange with you where you said we shouldn't put any powerful items in the game because the AI would never be able to handle it, so it sounds like you have very high standards for AI. If I give you a figure and we reach that figure, there's no guarantee the AI we produce will live up to your standards even if I and the AI programmer think we've come up with something pretty good. Do you still want me to give you a figure?

Also, if I hide your post for unrelated reasons and I've not answered a question you want answered, please just repost the question without all the extra commentary. Unsurprisingly, I've been really busy for the last couple of days - that doesn't mean I'm deliberately avoiding answering your questions.

Share this post


Link to post
Share on other sites
14 minutes ago, Chris said:

Do you still want me to give you a figure?

Yes. In pretty much any game review, the critic will comment on whether the graphics are good or not so much, so there's pretty obviously an industry standard of what is considered "good graphics", and same goes for AI. If you want a testable metric: if you give the AI the same units that the player has, and your test players react with "That was tough" or "That was a challenge", then you almost certainly have an AI that is good because of tactical / strategic decisions rather than number superiourity.

Share this post


Link to post
Share on other sites

Alright. In my head hitting £50,000 gives us some decent breathing space and at £100,000+ we should be able to sit our AI programmer down for a month / six weeks and let him work on the "advanced" AI without worrying about other features.

To me, that should be enough to produce a challenging AI. If we only get £50,000, I still expect the AI to be better than in Xenonauts 1 but I'm not sure we'll have the luxury of giving our technical lead a long uninterrupted period to work on the AI alone (which is probably what we need for something genuinely good).

Please remember all the disclaimers I gave in my first post.

Share this post


Link to post
Share on other sites

Why, why in any possible world, would you want the program fine-tuned for a situation that isn't the actual final game?

An algorithm that is good with units that aren't the ones it will be controlling, or good against units that aren't the ones it will be fighting, is unlikely to be good at its actual job.

A fun system will end up accentuating the intended strengths of its units and not  negate the intended weaknesses except in a way that is appropriate to the units being simulated. A reaper aware of an enemy far across an open field will run towards it, because that's what reapers do, even though a better tactic would be to hide in a blind corner until within or behind enemy lines and attack suppressed troops from behind.

Six weeks straight of programming might not be the most efficient use of time/money. A few iterations of candidates sent to beta testers with the instruction "Find the cheap strategies that are way too effective", followed by algorithm or combat balance adjustments that make those strategies not excessively effective is probably the way to go. 

On 6/20/2018 at 11:56 AM, Drakon said:

Yes. In pretty much any game review, the critic will comment on whether the graphics are good or not so much, so there's pretty obviously an industry standard of what is considered "good graphics", and same goes for AI. If you want a testable metric: if you give the AI the same units that the player has, and your test players react with "That was tough" or "That was a challenge", then you almost certainly have an AI that is good because of tactical / strategic decisions rather than number superiourity.

 

Share this post


Link to post
Share on other sites
On 6/24/2018 at 8:41 AM, Decius said:

An algorithm that is good with units that aren't the ones it will be controlling, or good against units that aren't the ones it will be fighting, is unlikely to be good at its actual job.

I'm sorry, but as a programmer i can tell you that you are plain wrong. The basic steps for pretty much any unit in Xenonauts are virtually identical: move to optimal position, engage target if possible, repeat until no further effective moves can be made with the remaining TU. In the vanilla non-CE Xenonauts 1 i would even expect that all single tile ai-controllable units excluding reapers were using literally the same algorithms, with optional sub-routines for special abilities like flight or teleportation.

Please keep in mind that the original topic of discussion was primarily whether the question "Is this an AI that is challenging due to strategic/tactical decisions rather than just stat / number superiourity?" can even be reasonably objectively answered.

 

On 6/24/2018 at 8:41 AM, Decius said:

A reaper aware of an enemy far across an open field will run towards it, because that's what reapers do, even though a better tactic would be to hide in a blind corner until within or behind enemy lines and attack suppressed troops from behind.

Actually, you just wrote a good scenario of how reapers could be shifted from slightly annoying nuisances to the actual terror troops they are supposed to represent. I tell you if reapers acted that smart, even if their stats were sub par, players would break out in cold sweat as soon as they hear that reapers might be in the operation area. And that will get you WAY better player reviews than just a super tanky no brain reaper that runs straight at the player, where it is not a skill challenge of positioning and identifying threat areas but instead a test on if you brought enough equipment to be able to gun down the reaper before it reaches your troops.

 

On 6/24/2018 at 8:41 AM, Decius said:

Six weeks straight of programming might not be the most efficient use of time/money. A few iterations of candidates sent to beta testers with the instruction "Find the cheap strategies that are way too effective", followed by algorithm or combat balance adjustments that make those strategies not excessively effective is probably the way to go.

Again, you might want to leave these things to people who are actual programmers. The fine tuning you speak of is definitely good and necessary, but implementing things like basic AI squad tactics is something that needs to be done before you can even send out your prototype to the testers for them to evaluate whether what you consider squad tactics is even worth anything ... and writing that AI code costs time and money.

  • Like 1

Share this post


Link to post
Share on other sites

Should an advanced AI be able to make errors? What I mean is, in case the programmers get to a "level 5" AI in the end, could a single unit have a chance to fumble, especially when under stress, and behave at "level 3"? History is full of bad decisions taken for various reasons. Does "fuzzy" AI exist, where you replace each branching by a set of branching, depending on (directed) random fortune?

For example, the duel turned to be a draw and both the alien and you are half dead. It's the alien turn and "he" knows that if he stays here one turn more, he will be dead, plus he sees that your distant mates prepare to shoot at him. Based on the AI (mainly the distance of your mates), and the designed alien psychology, the alien would either shoot, kill you, and be soon dead, or take the chance and run for cover, because he thinks that he can still do more damage to more people, on the next turns. This kind of decision has not to be deterministic, except for (terran only?) robotic units, mind controled units or cheap clones that have little concern for themselves. I'd love a game where the alien would shoot at you (as you expect), but only in 6 cases out of 8, and would take the chance in 2 cases out of 8. Another example is when an omniscient enemy AI would have the map revealed and two of your snipers located, and still have one chance on 10 to send a unit in the open just in front of your snipers.

Obviously, with such a "fuzzy" advanced AI, the concept of confusion would shine: just unbalance the finely tuned set of random chances, and make the most viable behaviour the less likely to occur.

  • Like 1

Share this post


Link to post
Share on other sites

Smartest, most efficient and logical AI isn't always the one that offers the most fun game play experience. Sure, it should offer a hard enough challenge, but I feel that a "good AI" requires a lot more than an opponent that does all the right decisions in every combat scenario. AI should have varied, fun and unexpected reactions to players actions so that a player could never be 100% sure what an opponent does, even if the "right action" was obvious. AI can and should do wrong moves every now and then. Stupid mistakes are fun, as long as there's some logic why the enemy does what it does. Moving back and forth between two covers burning all the TU's and not run away or shoot a Xenonaut standing in the adjacent tile is wrong kind of stupid behavior. Don't want that. :)

Edited by Skitso
  • Like 1

Share this post


Link to post
Share on other sites

One of the weird AI behaviour in AI War is that when the AI think it's going to lost a combat it'll commands its units to evenly split to all available escape points.  This is non-optimal from a human's point of view, but it is designed to prevent players from exploiting AI behaviour and funnel them into specific escape point(s).  So what you think is smart or stupid may not be actually that smart or stupid when applied to AI.

It's one of the very, very few games that has a gradual AI difficulty setting, possibly the only one in modern time, from 1 to 10, and the differences is very big.

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
Sign in to follow this  

×