The sorian PA AI thread

i can understand unit cap for online games but for offline games there shouldnt be a limit.

Really afraid of our taking up all the server resources with its agressiveness?

As a precaution for when the AI becomes sentient and tries to take over the world?

I've noticed that when left alone on a planet, the AI will tend to fill up its available population with dox, and not much else.

Does it know to destroy its units to make room for more appropriate units for defending a planet/inavading via teles? Which it actually does very well I've encountered.

AI PTE Report (like the one before 77965 where apperently the AI makes better attack choices...have to check that out)
~~Bad Stuff
Ok, im still going to hit AI on orbital, because it desperately needs improvement.
Here we have 3 separate AI on a gas giant, building 2-3 jigz each...(ya thats it, even though eco suggests they build more) and they start building 5-6 orbital factories, using only 1 anyways, and building NO avengers nor Fabbers, and the units they do build (SXX and Radar) that can be sent to different planets are completely forgotten and jumbled together....with other AI stuff...which makes for very weird traffic jams...
observe

AI still sucks at invading, why is it whenever it lands a fabber it decides to build mexes or power plants? and even when they build teleporters they will often not link the teleporters for a while and it ends up getting destroyed when they could have been streaming in units...

~~Good Stuff
i read that you made it so AI will never use orbital transports for commanders, thats brilliant, ill take your word on the fact that it works, and i approve.
AI is much faster at expanding now, though i still think it can be greedy if it finds itself alone on a planet, and take a lot more mex
Really the good points of AI is its great at single planet brawling, it makes good choices with attack avenues and uses eco and builds many factories to create larger armies. Its when it gets to multi-systems that it becomes very weird (understandable because no other RTS actually playes on several maps...at once...let alone round maps)

~~Other Stuff
AI loves building nukes and using them with....downright cruel....results
AI however even when its been nuked is very slow to build anti-nukes or really focus on getting them up fast.

For the SXX and radar crowding, I've noticed the AI likes to conserve most of it's units in a big circlejerk near a factory until it thinks it can get a snipe or massive damage to the enemy base, so it might be waiting until it gets a certain number of SXX to thrown them at the enemy commander all at once

- Unfortunately the ai still likes to build a vehicle factory on the small island of the 1vs1 pacific map.

- I bet that this is a very tricky point, but the ai is very vulnerable to snipes (booms, infernos, sxx and t1 bombers). Is there any way that it could detect an incoming snipe and temporary reacts with pulling everything they could out to prevent it.

Like if there are one of the following units massed and detected by the scout:

boom bots -> some dox patrol around the com
infernos -> all forces close by are retreating to the com
t1 bombers -> aa patrols around com, com build a few aa turrets

if sxx detected:
sxx -> all fabbers around + com build an umbrella (if no orbital) if orbital some avengers patrol next to the com

if unit cannon or nuke detected:
com stays close to umbrellas and anti nukes.

- I really like how the ai is able to rebuild after parts of it's base got destroyed and how it tries to attack always in the stop which is less defended.

that is heavily disappointing, I do not like unit caps, do you think there is a chance that this decision may be reversed?

Seems the audio has been blocked by Twitch on your stream recording, Sorian. Looking forward to that YouTube upload...

For those who missed the topic (like me)

Now we have subs will you be introducing a threat variable for them? I really want to be able to do Sub to AntiSub threat checks in the same way Air to AntiAir is done.

Hey just a question I was curious about. When issueing commands does the ai manually click everything it wants i.e individual placement of mex and attack commands or does it use area commands purely or does it do a mixture of both?

Hey @sorian , I'm watching your stream and you say that attacking the commander is a hard scenario to learn because it happens so infrequently in training.... but if you rewarded the AI for causing damage to the commander (instead of just rewarding them for killing the com), that would be much more likely to happen in training and make the AI learn that reducing the commander's health, even if the cost is high, is valuable. Even if the AI causes some damage to the commander (say, 25% reduction in HP), it would only take 4 attacks like that to win the game. Is it possible for your training to reinforce this type of knowledge?

I'm running through the latest AI changes and I think Orbital Launcher - Rush in fabber_orbital_builds can be purged. It's designed to be for those instances where the AI wants to get off planet with its commander NOW, except that the section for giving the TransportToPlanet order in command_misc_builds was purged. Therefore the AI will rush this launcher to get the commander off-planet but never actually send the commander off planet using a transport from the launcher.

Maybe, rather than purging Orbital Launcher - Rush, it'd be worth bringing back Move Commander To Planet, but making it lower priority than a teleporter and tying both it and the rush into NoWhereToHide so it's a really last ditch thing. That's what I'm going to try

It makes sense to me. Spinners are built according to current Air to AntiAir threat, while Narwhals should be built according to that and Sub threat because these are the things they counter.

While I'm posting wishes, it'd be great to have the AI have some way of knowing the number of players and how many starting planets existed. That way I could potentially make more intelligent decisions about what to do based on the radius of the planet vs. the number of players (don't go Dox on a 10 player FFA on a 250 radius planet). A measure of base to nearest enemy base threat would be great too, though I imagine much harder.

So you're saying that when the AI decides yes/no to attack, it weighs the metal-value of the damage it can cause versus the metal-value of the damage it is likely to receive? Or does it compare the number of units potentially destroyed on both sides? What is the metric or heuristic?

In every multi planet game I play the strategy of the AI is always nuke nuke nuke nuke nuke. So in order to win, all you need to do is cover your planet with anti nukes. Invade their planet, and then you will win. Is this intentional, cuz honestly its quite boring.