Making the AI smarter

What are some things the AI could do that would make life difficult for players? Not “f2 attack” the second the game starts obviously, but more like little micro tricks in combat that a moderately aware human player would do that the AI doesn’t.

Some I can think of:

  • targeting carriers
  • kiting with fast units like vultures
  • targeting fragile units like casters or workers
  • retreating from bad fights (to later stack waves)

I’m not especially tech savvy so I don’t know how hard or easy any of this would be to implement, but the AI in coop is extremely rote. It would change the game a lot to give it some tactics.

1 Like

That would make the game overly frustrating to play and wouldn’t be fun as it would make some things useless as well as making mutators way more aggravating than they already are. If I wanted to deal with that, I’d go play VS mode.

1 Like

I’m not saying it’s a good idea! Just tactically obvious things the AI could be doing to make player’s lives difficult.

1 Like

Why for? They are Amon’s puppets. It would be strange they display any advanced tactics on the battlefield.

Yeah but they’re not automatons either though, presumably. Some level of “we’re not ants” strategy would make sense.

Just having any RNG for attack waves would make missions more interesting. A lot easier to implement than improving the AI or putting in specific strategies/cheese.

For example, yesterday, I played Temple of the Past with a random Zagara ally who had clearly memorised all the spawn waves’ locations and timings (https://www.starcraft2coop.com/missions/templeofthepast#timings). Moved their ling & bane army to the spawn locations beforehand and that was it. I barely got to do any killing, let alone needing to even try.

If the attack waves were randomised, none of that camping or prep would be possible. (I’m aware that the SC2 engine is deterministic but just like enemy compositions, certain things can be randomised on initState or lobby.)

4 Likes

Imagine if that Deepmind AI was somehow ported into SC2 coop as a separate difficulty :slight_smile:

No lame mutators, just super smart AI.

Yes. It’s called AlphaStar. https://deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-StarCraft-II-using-multi-agent-reinforcement-learning

Pretty sure most of the “git gud bruh” ppl here would insta quit. If they were good enough to handle AlphaStar, they’d be in the GM league and not spending time in Co-op.

More importantly, Amon’s forces get a head start on all missions. Put that in the hands of even a noob player, let alone a good AI, and the Amon player can wipe the 2 Co-op allies. If AlphaStar’s “best” is plugged in for Brutal mode, people would just stop playing Brutal.

I wouldn’t mind Dark Templar marching into your base and a-moving whatever they first run into, and similar things like that. I think if you want to make it any more complex than that, you need to tie it to Amon having actual resources and income that can be disrupted.

If they did a “Team Coop” mode, where team Amon is actually controlled by human players, this would be the sort of thing they could do.

Right off the bat:
–Take out singleton Pylons powering a whole group of Protoss buildings
You could really wreck Artanis since he especially tends to do that.
You could even go so far as to evaluate that having only 2 Pylons, taking them out is still the more optimal way vs. taking out 8 cannons

–take out your lone detector
Then send in cloaked/burrowed units

–take out your high DPS units
If the objective is to destroy some unit or building with 10K to 30K hp and optional shields, then go after Immortals or Ascendants

–take out high investment assets
So Ascendants with 10-stacks of Supplicant sacrifices. Prioritize lv3 Royal Guard vets. Ignore expendable threats (e.g. Stukov’s infested walkers, or Zagara’s units) and go after the units they’re protecting (e.g. Inf. Db’s, or Nova’s Units

–take out other key assets
For example, if the comp prioritized Rattlesnake’s revitalizer, he loses in the short term, but may come out ahead long term if a sustained attack forces him to not have enough them due to cd and being out of charges

–worker drops
DTs, Reavers, HTs, Lurkers, or Tanks that are far out, but still enough to hit your workers in Siege range

–worker drops Xtreme
They don’t even need transport ships. They could drop pod or warp in. Arbiter Recall, eat your heart out :kiss:

–mucho nuking
I was told Blizzard scaled this back due to just that… the community complained it was too frustrating

If we were to do a thing like “Team Coop” like I suggested above, there would need to be massive balancing for sure.

Another approach would be to not let the Amon player(s) build stuff, but just give him intermittent waves to control. He has no control over base production, and buildings already in place.

I played with a Karax ally who knew the spawn points for all of the Void Thrashers on TotP (there’s a certain spot where one grouping of towers covers both spawn points on a given direction). They were able to handle them just fine. I didn’t even need to send any forces to push!

Ditto for other missions where they know exactly where to set up towers to intercept all attack waves.

For a mission like Void Launch. They clear the enemy encampments, and then just set up towers and/or attack forces around all 3 shuttle start points. You can pretty much just go watch Netflix at this point :stuck_out_tongue: Don’t get me wrong, this is fun too. Especially when you do it on a Brutation (a normally damn difficult ordeal, but you lots of XP and achievement progress to boot). However, some of the games where we barely win, or even outright lost, have been more fun, and definitely more memorable to break up that monotony.

I’ve proposed procedurally generated missions, in part or in full. So this is what happens in the Diablo games (other games like Dead Cells and A Robot Named Fight also have this). Someone pointed out that some missions already have this to some degree which is true, but it is woefully inadequate.

I’m thinking stuff like for Oblivious Express, change the direction and trajectory of the train tracks. However, the Coop Developer diary/notes (on the fandom) states that when they made new missions, a lot of thought and planning goes into making sure there’s enough space for both players, enough space in expansions, how far out they are, enemy spawn points, objectives, and other balance factors that I’m not sure if they could come up with an algorithm to deal with all that. Worse come to worse, just “suck it up” in where some “seeds” may not be fair, or too difficult for certain comps and COs.

Shame, gave a small but real change to gameplay when you had a comp that’d nuke.

They already have a smarter AI programmed. It’s used when you play VS AI games in Versus mode. It micros its units and does things like blinking stalkers to avoid AoE. They don’t need to use machine-learning AIs.

Co-op mode uses the campaign AI. The campaign AI is deliberately designed to be dumb. Blizzard themselves said as much a million years ago: http://classic.battle.net/scc/faq/aiscripts.shtml

Campaign AIs are meant to be beaten.

If you think the AI needs to be “smarter,” you don’t completely understand the design of the game mode.

The proper way to add more difficulty is through randomization/unpredictability or by splitting the player’s attention more (i.e., forcing the player to do multiple things at once, maybe on different parts of the map, which is how the bonus objectives are often designed). Or mutators. Or people could just accept that they got really super good at a game mode that’s designed to be relatively easy and beatable and play something else if they want to tryhard.

The AI does seem to have some intelligence in what it does, beyond simple A-moving. Amon’s Cyclones and Diamondbacks will kite you when they can, and their spellcasters have incredible predicting abilities in targeting your forces. They also decent priority in targeting hurt/expensive units in your army.

2 Likes

Nukes were pretty the main reason why any CO ended up making towers, no matter how awful they were… to fend off Ghosts.

After getting nuked numerous times, and eventually playing all of the COs well into Mastery, one of the things that entered my mind in addition to the usual slew of questions* was “what options does each CO have to deal with nukes?”

*. what are my AA options? DPS? AG (anti-ground)? panic buttons? healing? support units? what things help my ally?mineral dump? etc.

Blizzard has tried to make the AI smarter in the past, but many of those changes had been scaled back if not reverted. For example, there was a time in co-op when the AI would deliberately target carriers over the interceptors, which (at the time) made Fenix virtually unplayable. (This was a time when interceptors cost money and conservators were garbage, making mass carriers Fenix’s go-to build.)

I do think that people are pining for an AI that is meant to be beaten (it wouldn’t be co-op otherwise - no one wants an AI that rushes the players with all their pre-placed units the moment the game starts), but still hold its own compared to ladder AI without the benefit of handicaps (ie. mutators). That’s a very difficult balance to achieve, since a tweak of the numbers here and there may make some commanders better than others. (For example, simply giving the enemy more units will unfairly advantage commanders like Dehaka, who can turn that around and ramp up faster.)

Others are suggesting that instead of making smarter AI, it may be better to tweak the balance of the maps themselves, giving enemies more varied spawn points or timings, changing the composition of pre-placed troops, and so on. It’s been noted that Scythe of Amon is generally harder than, for example, Rifts to Korhal (cheese notwithstanding), and part of that is the pre-placed troops and the mechanics of its objectives. Indeed, most co-op balance changes that don’t involve commanders are generally about mutator and map interactions.

Given developer investment, however, it does seem that if you want a challenge but want it in a controlled environment (more controlled than Brutal+), they are steering you towards custom mutations. That’s not a good or bad thing, but I do feel they need to incentivize custom mutation play by offering some form of EXP boost relative to difficulty of the chosen mutators, even if it is less than that of drawing the same mutators in Brutal+.

Yes, this is exactly how I imagine it should be if it were to work in a reasonably balanced way. Including having zones where they can’t attack until a certain time. Eg. 1st attack wave to base, 2nd to expo or main, 3rd to army (not expo or main), etc. and then some combos. And otherwise having waves attack anywhere else too. (Need to prevent “parking” attack waves to stack them with the next one, since Amon’s team will have better economy and ready prod buildings.)

I think spawn camping the shuttles is okay, just like spawn camping the trains on Oblivion. If there were separate and random attack waves headed to the bases, then players would need to choose to split armies or build static defense, etc. (However, changing the direction for some objectives in both missions would make it more interesting, regardless of randomising attack waves.)


These discussions lead me to believe that the simplest (and stupidest) solution is to perma-activate Aggressive Deployment on Brutal - but even that has known timings. (Though, with unknown but deterministic drop locations.)

You are not the first to suggest it. I have been suggesting it back in 2018.

With advanced AI like,

_ Mutas clump together and target 1 unit each time to ensure 1-volley kill
_ Muta, Zerglings will automatically spread out when meet banelings or avoid Siege Tank shot
_ Transport ignore turret attack and went straight to the mineral line

Advanced AI should be a better choice for brutal+ than the bs Custom Mutation we are having.

The AI is quite good compare to other games, AI macro is solid, it is more than enough to train you up to bronze, or even silver.

Vultures? 20 characters