AI Roguelite

AI Roguelite

Character Autonomy 2.0 (Self-driven NPC Behaviour and Dialogue)
27 条留言
LeakyCranium 9 月 12 日 上午 1:06 
I would *love* the modified sentiment check.
Nathan Pascal  [作者] 9 月 10 日 下午 3:56 
Oh, another topic.
For personal use I run a modified sentiment check which is much more restrictive, so you get relationship changes less often (almost never). A weary glance does not turn someone into your enemy, a nice word does not turn someone into a devoted friend, relationships remain neutral much longer and only develop over the duration of a campaign, normal business and everyday interactions will usually produce no changes at all.

Is that something people would want me to include here? Or do you like the current system, which is more... responsive/volatile?
Nathan Pascal  [作者] 9 月 9 日 下午 4:04 
Alright, combat changes are out for now, until I have (way) more time to test and investigate. Sorry for the... inconvenience? I guess, it was too easy, so... sorry for the convencience? No, that does not sound right either.

At least the main part seems to work fine and as expected for everyone, that is good to hear.
LeakyCranium 9 月 9 日 下午 1:34 
I've done most of my testing for this at level 1 as a gladiator. I tend to use tier 3 opponents, but have had the same results against tier 4. The character class I'm using describes character as experienced, but only have one skill and a weapon to assist my roles and my opponents are usually also lore wise experienced combatants, so I should at best be equal in strength to the opponents yes. I've also had problems though with clearly superior opponents, such as larger foes or enhanced creatures, getting wildly outperformed as well.
Nathan Pascal  [作者] 9 月 9 日 下午 1:12 
Would you say, you and your opponents are equal in strength lore-wise (what the descriptions imply), or are you playing a character that is implied to be masterful against more regular enemies?

Trying to figure out what might happen there, because in my own testing as a generic young adventurer I was hit pretty hard by enemies, but the new prompts might be biased to favor your narrative role over game mechanics a bit more.

But when this happens as a regular fighter against (lore-wise) stronger opponents, I have little idea what might cause it. Maybe I need to remove the combar instructions after all, if they behave so strangely for some.
LeakyCranium 9 月 9 日 下午 1:03 
Oh I'll note that things are a good bit more appropriate on Insane difficulty, though even then enemies tend to take fat damage nearly every turn, though at least with that setting ticked on I have enough critical failures where it evens out a bit more (though ideally i would like to avoid leaving it on for non combat reasons).
LeakyCranium 9 月 9 日 下午 1:01 
Its an interesting thing: what often happens is my failures are "half" failures. I will graze an opponent or fail to dodge, but somehow still get a traded blow in or something. With the mod active, the ai essentially always wants me to be putting out some damage, and occasionally npc's wounds will also just cleave huge amounts of hp off. This is on immersive for the record.
Nathan Pascal  [作者] 9 月 9 日 上午 11:05 
Mm, alright, I need to look into that more, it seems. Thanks for the feedback. (Although every combat maneuver succeeding sounds unusual, those would be governed by the player roll checks mostly.)
LeakyCranium 9 月 9 日 上午 9:45 
Seconding Reffy's feedback. Nearly every combat maneuver succeeds, and enemies struggle to do anything to me. When I do take damage, its very little damage. Also seconding his input on the narration: everything else is great, but for some reason the AI interprets the combat prompts in the mod as "make everything easier" rather than "add texture and complexity to combat"
Reffy 9 月 3 日 下午 2:59 
No damage numbers. I meant that I don't get damaged at all. The red knife icon doesn't appear. Even though the text says I got hit.

Again, might just be the difficulty of the base game being weird.
Nathan Pascal  [作者] 9 月 3 日 下午 2:11 
Mm, when you mean damage numbers, those I have not changed here.
If you mean, enemies never try (or succeed) to hit you, that might be because of the mod, but relatively situational.

Regardless, happy to hear it could solve many of the other narration problems, those I spent way more time on to hopefully get right.
Reffy 9 月 3 日 下午 1:53 
The difficulty; Adventure, Roguelite and Insane; I think it's a bit scuffed.
Insane I get so much damage, but on Roguelite and Adventure, I rarely if at all I get damaged. ?
So might not be a problem of this mod.
Reffy 9 月 3 日 下午 1:26 
This mod fixed a lot of problems I had with the base game.
NPCs are more interesting, and I don't get the "but then, more enemies approach!" infinite mess that the AI usually throws at me.

The combat might be a bit too easy in this mod. I rarely get damaged, regardless what happens in combat.

Other than that, good mod.
Nathan Pascal  [作者] 8 月 30 日 下午 1:45 
Glad to hear.
I was a little concerned, I might have overdone the pacing. I pushed pretty hard towards momentum and resolution, since in my old games (before 2.0) NPCs were always stuck in choice paralysis and could rarely make up their mind to take any action at all.

So I am a little surprised, it has the opposite effect for you, slowing down the pacing, but I suppose, all taken together the instructions balance out nicely like that.

Mysteries of the AI, I am sure, the next model update will bring new challenges.
But for now, thanks for the feedback!
LeakyCranium 8 月 30 日 下午 1:02 
Just wanted to say your mod is fantastic. I was having issues with npcs randomly appearing in scenes they werent warranted in and rushed pacing, but youve nailed a much more even handed pacing that still allows for very fluid conversations. I slap this on top of reactive realms, and letting this override seems to result in the game working as a generally flawless simulator. Thanks for this one (also love the epic quests mod)
Nathan Pascal  [作者] 8 月 23 日 下午 8:04 
2.0 is here.
Happy days.
I hope, the new changes are welcome (and nothing strange is broken now, the AI always finds a way to be weird, I feel.)
Nathan Pascal  [作者] 7 月 25 日 下午 8:09 
Well, for some reason this mod is trending again, and I am still using it regularly without any problems, so I have overcome my Plausibility 2.0 consternation and reworked the entire prompt again.

Purging all hesitation and vagueness from the NPCs as much as possible is my new goal. While realistic, I eventually found it a little tiring, so now I test a new version where everyone is more decisive and pro-active. Wish me luck. (Not yet updated, just teasing.)
Nathan Pascal  [作者] 7 月 9 日 上午 2:39 
The two mods overlap at some points, so probably not very compatible, I fear.
Kamiyama 7 月 9 日 上午 2:33 
Is this compatible with reactive realms or would the two interfere with each other?
Nathan Pascal  [作者] 6 月 21 日 下午 10:08 
Experience, certainly.
I sat with my character in a teahouse over dozens of incarnations, with different levels of NPC behaviour, I had an ensemble of monks dancing around me, waving silk robes into my face while musing about the transcience of time; I watched an old monk clean the teahouse for 70 turns, while nothing ever happened; I sat in silence while NPCs begged me to speak until I sang.

Still working on some other mods, the quest system haunting me, mostly. Never quite right.

And if nothing else, I still have my changes for my own games, even if a little troublesome with the new implementation, it mostly works fine for me.
Recipient Lumber 6 月 21 日 下午 4:10 
Im sure you probably maybe gained something from this. Experience? Reputation?
aughtfour 5 月 2 日 下午 2:17 
I too had taken pains to make the NPCs think for themselves, but nowhere near your efforts. And other improvements I've made over the past year have eventually become superseded by updates superior to or more sweeping than mine. I hope you'll continue to make thoughtful ed/additions to the game, especially now that you're even smarter.
Nathan Pascal  [作者] 4 月 25 日 上午 11:04 
So, the Plausibility 2.0 update makes this kinda awkward. We lost the individual talk-to-NPC prompts, now every action goes through the same set of fail/success prompts. Placing the new dialogue instructions there works, but is not optimal for compatibility and token efficiency.

On the other hand, the updated models (especially free, which I am using) seem to handle NPC dialogue much better without additional guidance, making the difference between base and my mod barely noticeable for me, so I am unsure how to support this project in the future.

Continue_without_input still works as normal, to fill the scene with NPC momentum, which is the main feature for now, it seems. Mm.
Nathan Pascal  [作者] 4 月 14 日 下午 7:15 
And another, another update after testing. NPCs now notice you a little more even when silent, nobody dragged Elara or Elias into the scene, although it might still happen, I guess. Cannot get the Elara out of the AI. For best no_input experience I still recommend de-activating the quest tracker in calm scenes, unless you want NPCs to engage with your quest (which will make sense in some scenarios, of course.)

Balancing all the different AI tendencies is surprisingly difficult, but this is it now, I hope. Already was scratching the token limit, trying to persuade internal weighting. Output varies between scenes quite a lot, so evaluating averages is always hard, let me know if you encounter some persistent, obsessive behaviour from the AI (that is not the usual ones, I mean. If I see one more crack in the walls opening with black ooze I will burn that house down.)
Nathan Pascal  [作者] 4 月 14 日 上午 6:06 
And another small update, clearing up a bit, dancing around the token limit carefully and fearfully.

One thing I still dislike, the AI does not completely stop to assign names to characters in no_input, which must be stopped, or Elias and Elara will take over all possible worlds by themselves eventually.
Nathan Pascal  [作者] 4 月 14 日 上午 2:27 
Another small update, without input the NPCs got a little too excited about your current quest task when active, wanting to solve it themselves, which is a bit too much agency. This behaviour - while not eliminated - should now be reduced.
Nathan Pascal  [作者] 4 月 13 日 上午 5:01 
I have an update for you. For one; I had put the story_preamble in the wrong folder so it did nothing until now. Well... on another note, I spend another dozen hours to optimize the prompts and, honestly, I am breath-taken what you can get out of the free-tier system with enough coaching.

I believe this is all I can do - all that can be done, in this specific scope. The without_input option is heavily geared towards creating meaningful engagement in low-tension situations, so there might be some strange behaviour when done in high-stakes situations. I will try to observe and adjust if that is the case, but otherwise I am incredible happy with the results.

I hope you will be, too.