Things that AI needs to improve

Describe your experience with the latest version of FreeOrion to help us improve it.

Moderator: Oberlus

Forum rules
Always mention the exact version of FreeOrion you are testing.

When reporting an issue regarding the AI, if possible provide the relevant AI log file and a save game file that demonstrates the issue.
Post Reply
Message
Author
User avatar
Oberlus
Cosmic Dragon
Posts: 5715
Joined: Mon Apr 10, 2017 4:25 pm

Things that AI needs to improve

#1 Post by Oberlus »

Playing build 2019-04-29.1601105

- Suicidal routing:
I'm tired of watching AI lonely ships crashing into a well defended system trying to get to another system. To be more specific:
Empire A is reinforcing System 3. Two paths to get from System 1 to System 3:
Path a: System 1 (Empire A) <-> System 2 (Empire B) <-> System 3 (Empire A)
Path b: System 1 <-> System 4 (Empire A) <-> System 3
Some of the Empire A's ships choose the path a and are destroyed at system 2 causing no loses to Empire B.

- Forgotten troopers:
Each time an AI plans an invasion on a system, when the armed ships are destroyed (in their way to that system or when they arrive), the troop ships that followed the destroyed squadron stay put on wherever they are for the rest of the game (or until they are destroyed). This cripples AI production in two ways: wasting of resources on unused troop ships, and unnecessary extra ship upkeep.

User avatar
Oberlus
Cosmic Dragon
Posts: 5715
Joined: Mon Apr 10, 2017 4:25 pm

Re: Things that AI needs to improve

#2 Post by Oberlus »

- Unfair/cheating clairvoyance (this is only possible, not verified):
When two systems are 1 turn away (say distance <75 and all ships are speed 75+), and I have a stack in one system and the AI has a stack in the other, if I command my stack to attack the other one, AI knows it and retreats.
This does not happen always, but it happens more often than not when my stack is bigger than its stack. And it does not seem to be caused by the previous presence of my stack (i.e. retreating because my reinforcements arrived in the previous turn), since my stack can be there for many turns before I decide to attack (usually because I'm waiting to finish the troopers). Anyhow, the AI retreats in the same turn I send my attack, so it should not know I was attacking (the same I couldn't know if it was the AI who attacked me from its system).
Is the AI aware of my fleet orders when I hit next turn?

User avatar
The Silent One
Graphics
Posts: 1129
Joined: Tue Jul 01, 2003 8:27 pm

Re: Things that AI needs to improve

#3 Post by The Silent One »

Oberlus wrote: Tue May 07, 2019 2:56 pmIs the AI aware of my fleet orders when I hit next turn?
No. The AI clients get exactly the same information as the player. They can't cheat.
If I provided any images, code, scripts or other content here, it's released under GPL 2.0 and CC-BY-SA 3.0.

User avatar
Oberlus
Cosmic Dragon
Posts: 5715
Joined: Mon Apr 10, 2017 4:25 pm

Re: Things that AI needs to improve

#4 Post by Oberlus »

The Silent One wrote: Tue May 07, 2019 3:50 pmThe AI clients get exactly the same information as the player. They can't cheat.
Then that stuff I report there must be a fallacy based on coincidences.

Morlic
AI Contributor
Posts: 296
Joined: Tue Feb 17, 2015 11:54 am

Re: Things that AI needs to improve

#5 Post by Morlic »

Forum rules
Always mention the exact version of FreeOrion you are testing.

When reporting an issue regarding the AI, if possible provide the relevant AI log file and a save game file that demonstrates the issue.
Once again - the chance to fix the issues you mention increases by roughly infinity (i.e. zero to non-zero) if you could provide one or multiple savegames where one can actually see and debug the AI behavior and verify possible solutions.
If I provided any code, scripts or other content here, it's released under GPL 2.0 and CC-BY-SA 3.0

User avatar
Oberlus
Cosmic Dragon
Posts: 5715
Joined: Mon Apr 10, 2017 4:25 pm

Re: Things that AI needs to improve

#6 Post by Oberlus »

Morlic wrote: Tue May 07, 2019 6:36 pmOnce again - the chance to fix the issues you mention increases by roughly infinity (i.e. zero to non-zero) if you could provide one or multiple savegames where one can actually see and debug the AI behavior and verify possible solutions.
I'm sorry, Morlic, you are so right. Eventually, I'll open an issue on github with logs, saves, etc. for the problems I've stated. Here I was just opening this thread mostly to remember this and maybe to motivate others to tell about related stuff.

4xel
Space Floater
Posts: 16
Joined: Thu Apr 18, 2019 10:56 am

Re: Things that AI needs to improve

#7 Post by 4xel »

Oberlus wrote: Tue May 07, 2019 2:56 pm - Unfair/cheating clairvoyance (this is only possible, not verified)
I would be really surprised if it were anything more than a coincidence. If anything, AI does not retreat when it should, and it does invest resources in scouting.

Map reveal would be the mildest of the cheats IMO (scouting is quite hard, especially remembering what you saw and guesstimating how things might have changed, and even with map reveal, exploiting map knowledge is far from straightforward), though I'm really happy symmetry is a tenet of FO. By the way, is FO a WEGO?
Last edited by 4xel on Tue May 07, 2019 8:23 pm, edited 1 time in total.

User avatar
Oberlus
Cosmic Dragon
Posts: 5715
Joined: Mon Apr 10, 2017 4:25 pm

Re: Things that AI needs to improve

#8 Post by Oberlus »

4xel wrote: Tue May 07, 2019 8:19 pmI would be really surprised if it were anything more than a coincidence. If anything, AI does not retreat when it should, and it does invest resources in scouting.

Map reveal cheats would be the mildest one the cheat IMO, though I'm really happy symmetry is a tenet of FO.
Yes, it was probably some coincidences that caught my eye while I forgot about the (many?) times that AI don't retreat. Maybe it was going somewhere to do something else not related to my incoming attack.

Regarding scout, for the coincidences I have in mind, AI always had vision (systems are really close and having a colony nearby is enough to see up to three systems away).

slv
Space Floater
Posts: 26
Joined: Thu Mar 17, 2016 12:12 pm

Re: Things that AI needs to improve

#9 Post by slv »

Played a few games recently (after around 3 year break), I think I noticed few things which may (or may not) help AI. These aren't really bugs and it's somewhat opinionated but I think tehre are some tweaks which may make AI scarrier (I am overall really impressed by it, when playing other 4x games like stellaris or endless space AI is a total disaster, here it's much better)

First, I was checking AI tech sequence and it seems that they like to spread out too much. In my experience (and from reading other people's reactions) it seems that focusing on crucial gamechanging techs is a way to go (Force energy structures, N-dimensional structures, Xenological techs, Quantum networking, Distributed thought computing, Pure energy metabolism, N-dimensional assembly, Adaptive Automation, Stellar Tomography, Unified Consciousness, Enclave of the Void) and I found that trying to get your gamechanger technology as soon as possible is a good idea, so when I am playing I am trying to get one gamechanger technology, then get another one and keep opting for them without changing my priorities in the middle. I often see AI changing their mind too much so even if they generate approximately the same research by the time my Adaptive Automation or Quantum Networking finishes for me, for AIs they are only half way done they still keep switching between military technologies, growth technologies, pieces of adaptive automation. If instead they got all military technologies I would probably be much more threatened or if they had smaller amount of them and focused on a single thing (no matter which one, really) they wouldn't fall behind on economic technologies. Think AI should have higher focus on these gamechanging technologies I mentioned, choosing one of them and then sticking to it trying to get it as fast as possible without changing directions.

Speaking of military technologies, I found that AI researches way too much. They get military robotic hulls, asteroid hulls, organic hulls, massdrivers, lasers, shields and hangars. I feel like if they instead focused on only one (i.e. scattered multicellular as an universal and fast enough hull, speed is important for AI since it doesn't stack fleets too well) it would be much scarrier. Especially asteroid hull research seems to be a large research sink for AI (and for some reasons AI doesn't seem to research rock armor which is one of the main reasons to use that part of the research tree). A ship you can make using MD4, zortrium and military robotic or organic hull is of course worse than something AI makes, but not by much. Not wasting research on multiple hulls allows you to get more economic techs and win by sheer numbers, wouldn't be surprised if the same holds true for AI.

Similarly with weapons, while a ship with lasers is a big improvement over a ship with mass drivers and a ship with hangars is a good improvement over a ship with mass drivers, getting both seem to not give that much of a boost. At least in a current way the combat is handled a stack of 2weapon/2armor no shield ships is both simple to research (for AI), simple to produce and effective. It may be reasonable to ask multiplayer playtesters what they found to be good earlygame ship design and force AI to use that (I think I heard someone mentioning static multicellular hull as the go-to early hull).

Last, but not least, I think it is probably good idea to disable AI using defense focus. From what I saw it did more harm than good, sometimes even just a few ships made AI panic and switch half of the planets to defense focus. This does not hinder humans (human players know that it is just a temporary sertback and in the long war of attrition they can easily wait until they get enough of a doomstack to crush defense), it hurts only the AI which is very often overreacting. In fact from player perspective it makes the game much less challenging, the main thing which can offer some sort of a resistance is an AI who snowballed by killing all its neighbours. Disabling defense focus for AI would increase the chance of that happening, currently and AI which is beaten by another AI still dies, but significantly slows down the victor AI by just switching most of their planets to defense focus.

defaultuser
Juggernaut
Posts: 854
Joined: Wed Aug 26, 2015 6:15 pm

Re: Things that AI needs to improve

#10 Post by defaultuser »

Tech is highly dependent on the size of game and subsequent number of turns. Plus the focus of your main species.

Post Reply