[Deprecated] Considering the removal of AI Progress.

[Misery]

Okay, it's absolutely possible I'm missing something, or it's just the fact that I haven't been awake long and haven't had my caffeine, but one thing that kinda bothers me with temporary AI progress is, well, wouldn't it encourage the player to just sit there?  Waiting for that bit to go away before making another move?  That's something I've always noticed in gaming:  Give players ANY system that causes something to recharge/regenerate over time (which is basically what this is, they're waiting for "safety" to go back up), and they WILL abuse the hell out of it, and then whine that the game in question is slow/boring.  That they don't HAVE to use it that way never occurs to them, because doing it that way is the easier way out.

If the game design encourages players to do something, Devs don't get to complain when players do that thing. Even if that thing is not fun. Thats why WoW moved away from huge cool down abilities and using 5 flasks at once: players will do it once the game gets hard, to get every edge they can. Players will turn sit around waiting for their cool downs instead of playing, and complain that it's not fun, because making the raid even harder on themselves is foolish.

The absolute last thing AIW2 should do is add things that encourage waiting for the effect to go away. We do not need more Netflix time.

Yeah, pretty much.  I hate when games do any of that sort of thing.  Heck, regenerating health in FPS games... ugh.  That's one reason I don't play those anymore.

Definitely don't need more waiting for stuff in this.

[Steelpoint]

The general theme I'm getting here is that people are liking the idea of moving away from AIP into more a general overview of how the AI views the players actions in comparison to all the other things the AI is doing at the same time.

----------

A example being that on a scale of 100, the AI is devoting 20 units to the Spire war, 30 to Research & Development and 50 units to the as yet undisclosed major thing the AI is doing. With the player being worth not even 1 unit of attention to the AI.

However, as the player takes actions to the AI it will start draining attention from these other areas and start adding this to the player.

But by itself this system seems just a more complex version of AIP. But here's the catch in that the above system would give the devs the option of being able to better influence the attentions of the AI.

----------

For example, if the player were to somehow introduce a new threat to the AI, one that may even be just as hostile to the Humans, then that would also draw attention from the AI, thus draining attention away from the Humans. Or the player can do something to force the AI to put more attention to one of its pre-existing area's of interest.

----------

Another idea seems to be also to split the AI Territory into some kind of sub-sectors, with each sector having its own view of the player and other external problems. Only problem here is that unless you allow some carry over between sectors then if you kept all your expansion and attacks in one sector then none of the other sectors would ever go higher in alert status, unless there is a reason to raid other sectors.

----------

I feel that while AIP is a bit gamey, but it also is a clear metric of the players impact on the AI and how much attention you're drawing. Though I think currently AIP does not give much info as to what having a certain AIP level does aside from the tech up threshold.

[Elestan]

A key point worth pondering is "How is the AI's threat assessment fallible?".

This may seem like a subtle issue, but I think it sits at the center of the whole AIP concept.  The AI needs to ramp up its response to the player as the player does things that make them a greater threat, but if the AI can do this perfectly, you get a perpetual stalemate; every accomplishment the player makes is matched by a corresponding increase in AI capabilities.  If gaining a Golem gives the AI a Golem's worth of strength to fight it, then the player might as well ignore the Golem, because taking it gives them no relative advantage.

IMHO, what is needed is for the AI to be smart, but not all-knowing.  Guerrilla wars are about doing the unexpected, so let the player make gains that the AI doesn't know about, and give them ways to keep the AI from learning about a threat until it's too late. 

So I think we need a more sophisticated model of what the AI believes to be true, separate from actual ground truth.  We need systems, like Nebulas, or just Neutral (maybe resource-poor) systems, where the AI is not omnipresent, where the player can potentially gain or hide assets to pull out at strategic moments. 

I would love to be able to find a golem hidden in some backwater system, slowly repair it over time, move it carefully through space to a strategic position while avoiding AI patrols, then pull it out to use in a key battle, in which I use some kind of comm jammer and make sure no AI ships escape, to keep them from carrying word back to the Core.

TL;DR:  Being able to hide things is how you make it a guerrilla war, instead of a slugfest.

[NickAragua]

I like the idea of the AI having multiple "progress bars". I would suggest another alternative or supplemental system: have the AI  adjust its heuristics and priority weights based on player actions, kind of like a machine learning algorithm. So it'll be less likely over time to do things that are less effective and more likely over time to do things that are more effective.

For example, when the AI decides to send some ships to one of your systems to put its foot down, it has an "array" of ship types and odds of sending that particular ship type. It "rolls a die" and sends a wave of missile frigates against one of your systems, but you have an anti missile turret there and tractors, so they get exterminated. Now, for the next wave, the odds of the AI picking missile frigates against that system go down because that didn't really work. It might still send missile frigates, but it's more likely to send something else.

Etc. etc. for a bunch of different situations and situational properties. So you'd have an AI that adapts to counter the player's specific style, forcing the player to try other tactics and strategies in order to keep on top of things.

[Cyborg]

I like the idea of the AI having multiple "progress bars". I would suggest another alternative or supplemental system: have the AI  adjust its heuristics and priority weights based on player actions, kind of like a machine learning algorithm. So it'll be less likely over time to do things that are less effective and more likely over time to do things that are more effective.

There's a thread somewhere talking about this. The difference between a learning AI and an emergent AI. This is overly brief, but if I remember correctly, the benefits of the emergent AI are seeing actions performed in unexpected combinations that can challenge the player rather than a learning AI which is always recording.

A learning AI would be interesting for an expansion.

[Dominus Arbitrationis]

I think if you combined the two, you would have one of the most innovative strategy games of all time. And it would be quite possible for it to stand the test of time. Or, eventually learn more and more, decide that humans are unnecessary, and wipe us out. Either way, AI War!

Of course, that might be impossible, or just really really difficult to do, so I'm not holding my breath for it. But that would be a pretty amazing system. A learning, emergent AI. Sounds like something out of one of Asimov's books.

[Draco18s]

Just make sure that it's utility function prioritizes quantity of humans to play against, rather than crushing all human opposition.
We don't want to run out of milk to make ice cream.

[NickAragua]

Yeah, combining the two things would be pretty cool. You'd never want to completely eliminate the poorly performing options from the "bag of tricks", just reduce their incidence. And you'd probably want pretty good granularity, at least system-level. "This type of ship does not work well against this system". From what I've seen most other RTS games don't even do that basic type of learning, they just toss the same scripted squads against you ad nauseam. The idea, again, is to force the player to mix up their defensive setups and make the warfare more dynamic than in AIW:C, where territory is pretty static.

I guess it wouldn't make much of a difference if ship types for waves are decided randomly anyway - in that case, the AI gets what it gets and tries to help the incoming waves as best it can with whatever ships are available nearby.

As far as implementation difficulty, the learning part isn't really that difficult as an algorithm. As a first pass:
- you have a matrix of ship class x system, with each individual cell being the value of sending that type of ship into that system
- when the AI pulls together an assault fleet, it tries to maximize the value of ships being sent into the system. As a bonus, when estimating player strength in a system, use a 'last known' figure rather than perfect knowledge.
- the AI sends the fleet in (perhaps in support of a wave or some other event), and keeps track of each ship's "effectiveness", where "effectiveness" is some measure of damage delivered vs damage received
- ships with high effectiveness have their value for the system increased, ships with low effectiveness have their value for the system decreased

So the "learning" part is the construction of a fleet against a particular player system. The "emergent" part is what causes that fleet to be sent in - it might be in support of a wave, or in response to a power outage, or it simply becomes strong enough that the AI thinks it can overwhelm the local defenders. Or the AI needs reinforcements when a system comes under attack and then the nice shiny fleet gets sent into probably highly suboptimal circumstances.

The data collection will introduce a slight performance hit, since you have to record every attack made by/against a particular AI ship. However, the number of ship types and player systems is pretty low (tens), so running time and memory usage for the matrix update/fleet builder shouldn't be too bad at O(number of ships x number of systems), and it doesn't necessarily have to happen all that often.

You could also apply a similar learning mechanism to AI placement of defences, so if a player keeps using raid starships to pop command stations, you'll eventually find systems start to have a lot of orbital mass drivers and grav drills (whatever that thing is that slows everything down to crawling pace) because the other stuff proves ineffective at defence.

Sorry, I love ranting about AI and machine learning at a technical level, that's kind of my favorite comp sci problem.

[x4000]

Please note the results (for now) of this thread:

4.a. Temporarily Added AI Progress

and then the deprecated section:

3.a. Replacement of AI Progress

Thanks for the great discussion!  https://docs.google.com/document/d/1IdzU90psGas_3UFe23BLvsGQ8fclec49NmnbHfwkZ8w/edit#heading=h.f9cusllqwyxa