Skirmish AI: E323AI 3.22.4

[hoijui]

lagging can only cause a desync if there is a bug in the engine. but it is very possible that this bug gets a huge boost in likeliness to appear if there is lag.
In test games with AI (also E323AI), Zydox and me also came around desyncs. He even was able to supply some debugging data for it, but.. it is not yet fixed.

[MisterBenn]

I've seen when the commander assists the factory, that sometimes it jitters its torso to and fro playing the "order received sound" and never settles down to use its nanolathe properly. IIRC that didn't happen in 3.18.1 unless I happened to be lucky before.

Should I expect the AI to progress to T2 or is that still to be added in? I couldn't get a T2 factory to be made in either BA or SA when a bot played alone on a map. I haven't looked at the source, is there some combination of tags in the config files against factories or resource states that in theory make it possible?

Edit: After some more tests today I'd like to take a lot of that back. I decided to find out for sure with a 4 on 4 game, Arm versus Core, on Rusted Delta with all players 100% resource bonus. The results were mind blowing! There seems to be a bit of chance in getting to start the first T2 factory (i.e. be over the resource threshold in the config, have a builder free and choose the factory next) but with the options I chose, all the players got there.

I didn't see it to a conclusion as it made my E8400 cry in the end, but pretty much all the T2 land units were being put to good use. Many T2 tanks, Sumos, Cans, Pyros, and hundreds of T1 units in the mix. Not a lot of T2 defences though, but a couple of T1 plasmas were built. I presume that was just a matter of time as LLTs and Gaats get built quite often and the others are defined similarly in the configs.

There is much more to the AI than I realised from the first couple of tests! It seems that there's a bit of a hurdle to make the leap to T2, particularly on smaller maps with a lot of raiding but the bots can put together some great T2 attacks once they get there.

[Chamrin]

I can consistently reproduce the huge lag in the game with this AI. If I'm playing a flat land map, with a decently porked base, and I finish a huge value target, such as a nuke silo or advanced fusion reactor, immediately upon finishing the game slows to a crawl. Destroying/reclaiming it reverses the lag. I have a hunch that it's trying to find a safe path to raid it, and can't find one so it keeps trying and trying.

[MisterBenn]

I updated those config files to make this AI play Supreme Annihilation U41 once again. The changes are:

• Cleaned up the previous version that was produced by the BA config + the default SA template, and then scrambled a bit by my amateur "merge configs" shell script!
• Adjusted Crasher and Jethro to join attack groups in a long range role, rather than just be anti air (as was inherited from the BA config file.
• Went through every land unit and checked / adjusted their combat roles were effective!
• Made sure scout / harassment groups were effective (fast/armed/not a flea!).
• Made sure T1 attack groups contained a good mix of units (attack/artillery/aa) - for Bots/Vehicles/Core/Arm
• Small tweaks to the resource states

I ran many, many(!) tests today and the results with SA seem roughly as good as you can get from adjusting config files and I don't believe I've skewed the style of play away from how the bot plays other mods. It raids well as always and will throw relentless attack groups at you through all 3 tech levels.

After the testing today I see that the ability to progress to T2 is quite map and situation dependent. On large open maps like Comet Catcher, the AI needs to expand over quite an area and keep all the mexes in order to meet the expansion and resource requirements to tech up. Playing for example Small Divide, the mexes are closer concentrated and the bots tech up pretty reliably if the game does not get resolved quickly. I saw that resource boosts go a LONG way to helping. With 100% boost on Small Divide they hit T2 in about 7 minutes and after 25 you will have 2 Vanguards, a Razorback and a Bantha sat in your base!

Some feedback, Error323 if you are interested. I'll keep it to what I suspect are the more easily implemented :)

• The building of base defence structures is a good help but I suspect bases could do with more active defence sometimes. Would it be possible to use the DEFENSE tag against suitable units in the config so that a group or two could be built and remain at the base, and engage anyone causing damage?
• How about a SOLO tag to use against units to allow them to move at the enemy without being grouped first (or a group of 1)? This could be useful for Sharpshooters / Stealthy Spy bots and perhaps crawling bombs!
• Could slow reloading / high damage units be set to prioritise high value enemy targets? I've seen Penetrators finish off a half dead bot that's right next to a shiny new Sumo or Sentinel.
• What about a CLOAKABLE tag for units so that attack groups could be either entirely cloakable or not? You could easily whip up Gremlin / Seer groups to keep the enemy on his toes!
• There needs to be an ability for the bots to upgrade their existing mexes once they hit T2. New slots have mohos placed just fine but they also tend to be the slots at most risk!

I'm really impressed with the AI though and the configuration setup is good to work with!

[MisterBenn]

Hello again, here is my final version of the configs. I tweaked the tags against a couple of buildings and defenses and tweaked some of the tech 2 units further. I don't think I can take the config files any further and they're attached again.

If this AI is still being developed, I think the single biggest change that could be made is the prioritisation and choice of energy producers. All the AI games I have watched have had the players produce large sprawls of solar and wind generaters right through all tech levels. I have seen the more advanced energy makers get built (all the fusion reactors and the geothermals too, on the correct spots) but always very late in the game (around when tech 3 is appearing).

It seems the algorithm for choosing what to build is based upon a short build time as a priority. This is appropriate in many cases but with energy makers I think it is not so, where more expensive models are more cost and space efficient. I think the AI should be instructed to go ahead with the most expensive energy maker that it has available, that is within a certain build time (like perhaps 3 or 4 minutes for one builder). The AI should also not fear having a sizeable surplus of energy, as usage and metal makers are likely to catch up. I think this would go a long way to keeping the base sprawl in check too.

I final point: I struggled to get units to belong to both the scout groups and the main attack groups. The Peewee for example, is a great T1 scout and harassment bot, but it would be great to have it in attack groups too. As soon as the SCOUTER tag is used, the unit (unless I'm doing something wrong) will only ever be used in the scout groups and never in main attack groups, even if it also has ATTACKER and ASSAULT in there too. It would have been nice to be able to have them in both groups, as they are effective in main attack groups too.

[slogic]

I am not sure that time criteria is nice to make decision to build some advanced emaker. AI builds what is affordable by current time, sequentally scanning units from low cost to more expensive. This is general algorithm made by Error. I won't touch it by now. Also your suggestion may stall AI easily.

I wonder too like you about emakers, but about factories. They come too late between AI states 3..5. I think this is because algo which calculates affordable status returns false. Introducing another factory in early game may diverse AI's gameplay. This is more important for me than energy stuff which works well enough.

I know the problem about SCOUTER tag issue. All units with this tag are moved to SCOUT groups strictly. This is not easy to fix cause i need to keep and propagate unit tags from whishlist to military logic block. But i'm working on it.

[slogic]

More factories do not show up because of mNow/mStorage ratio, not function which returns affordable ability. It should be over 0.7 to let AI build second factory. I reduced it to 0.45, because AI is very good at resource consumption.

[slogic]

Error323 is absent so here is a prerelease.

Changelog:

• initial aircraft support
• universal algorithm of factory selection to be built
• fixed situation when units with SCOUTER tag can never be added to attack groups
• new scout units won't be requested if there are idle scouts already
• scout units won't be merged without real need
• ability to build defense for single building in a cluster
• tweaks to build second (etc.) factory earlier
• tweaks in target selection
• added config files for Supreme Annihilation (credits to MisterBenn)

Download (WIN32 ONLY)

[FaerieWithBoots]

maybe you should have a look at starcraft 2 for new version names Because what comes after high templar? XD

[JohannesH]

maybe you should have a look at starcraft 2 for new version names Because what comes after high templar? XD

arbiter

[Chamrin]

maybe you should have a look at starcraft 2 for new version names Because what comes after high templar? XD

arbiter

Archon is before Arbiter I think.

[MisterBenn]

FYI the SA config files weren't in that RC1 download. That's no bad thing though as for some reason they don't get on perfectly with 3.20 based on my tests last night, presumably down to some of the tags I adjusted when tuning for the previous version. I'll take a look as soon as I have a free moment and repost them.

[MisterBenn]

Here we go, the SA config files for E323AI 3.20 RC1.

I tweaked a great deal of the units in that old 3.19 SA config file to improve combat performance but the changes did not fare well under 3.20. It's clear that I should leave that stuff alone and leave the config files more "neutral" until while it's being so actively developed! These new configs are purely translated from the BA configs, with the following exception:

• Crasher and Jethro have SNIPER added to join land unit combat groups.
• Flea, Shellshocker and Wolverine disabled for same reasons as before.
• EMAKER added to Geothermal and Fusion power plants.

From a couple of tests, everything seems to be OK. Any reason why those power plants are disabled in the BA config files or is that just left over from the past? They seem to get used OK from what I saw.

BTW I also added my converter shell script, it reads a categorization file and adjusts unit and building tags within it where they exist in a second file and produces a merged file along with a list of entries that were not found. It automates taking a template file and using, for example, all the entries already in the BA config file where they exist. It's how I made the SA configs and could be used to churn out configs for all the other TA-like mods for Spring. It's attached on condition that noone derides my scripting skills!

Edit: it occurred to me with that script that I was working on files renamed to ba.cfg and sa.cfg - I never went back and put the bits in to make it cope with filenames containing spaces...

[slogic]

New release with the following differences from RC1:

• optimization when regrouping is needed
• fixed usage of improper threatmap for aircraft (surface threatmap was used)
• fixed a bug when group having at least one unit with ANTIAIR tag could stall forever
• tweaks in target selection so enemy units near defence matrix are boosted in priority
• brand new military group behaviour for bombers
• config files for Supreme Annihilation is really added in archive now :)
• tweaks for aircraft in BA 7.12 config file

Download (WIN32 ONLY)

[slogic]

...Any reason why those power plants are disabled in the BA config files or is that just left over from the past? They seem to get used OK from what I saw.

They are not really supported. They are working because of a side effect :). AI can't control the position at which GEO will be built.

[slogic]

For those who is unable to test new version here is a demo of the battle of 3.20.0RC2 vs 3.18.1 to see AI improvements.

Requirements to play a demo:
* Spring 0.81.2.1
* map Mount Dustmore
* mod/game BA 7.12

[Umrug]

* hugs slogic and MisterBenn *

[Error323]

What Umrug says! :D

[MisterBenn]

I'll accept a small proportion of those hugs, but I'll point out that I only contributed some config files and feedback - no actual development :)

Incidentally, I noticed a difference in how E323AI plays SA compared to BA. Where with BA the attack groups are sent regularly at the enemy, with SA they sometimes seem to collect and accumulate between the home and enemy bases. I have no idea why this happens when a great deal of the SA configs are the same as for BA but it does seem to be the case. Sometimes they sit building like mad and then with no apparent trigger suddenly send a mix of 150 leveler / raider / slasher etc at the enemy! It's quite amusing but probably doesn't pressure the opponent like regular attacks in BA.

[slogic]

Error323, i'll try to merge my branch into master tonight, so prepare some of your time to compile linux binary tomorrow. Geo is already supported in my branch. Also i have done tweaks to set separate categorization (not config) file per each team id to compare effectiveness of config files.

MisterBenn, this could be only if API's GetUnitPower() is not balanced for some units of SA mod. I mean for some unit it returns extra power value. AI will not send a NEW group of units if their cumulative power is less than target (with best score only = distance + threat - 150 * target_health_left_ratio) & also will merge existing groups of units until their power will be equal or more than target. This formula should be tweaked.

PS. Um, i lied a bit. For a new group minimal group size limit is used from config file. When group size reaches this value a group is sent to target whatever threat is.