the_hunger

First off, thanks for taking up where peppe left off.  I took a stab at it and found a weak solution.  You actually appear to have to fixed the crappy porting problem, so kuddos to you, sir. 

That said, I was doing some quick testing and just played a 1v1 vs nightmare ai.  As expected, it was easy for me to win, but ai is ai. But there were some obvious problems.  First, the positives, tp's worked great.  The ai even faked tps when I was working on its tower, which is behavior good players use, so that is perfect.  I didn't see any 5 foot tps, so that seems to be resolved and is a huge improvement over the peppe's last version. 

BUT - the ai is still terrible in a few areas.  In 1v1 on cataract - the ai had ZERO interest in flags.  It was more interested in attacking my towers in the middle.  The exact scenario was rook (me) vs reg (ai) on cata.  I held the flags 90% of the time, only retaking a flag like once or twice.  The rest, the ai just gave to me.  When I pushed a portal, the ai MAY have shown up a littler earlier than normal, but only half capped the portal flag back, and then never bothered to retake.  So, the bottom line, in 1v1, the ai is abysmal.  Giving up flags the entire game is a big problem, and the ai was much more concerned with hurting a few towers than retaking (even walking past flags it could easily kill just to start hitting towers).

So, +1 on tweaking peppe's tp code to make it work, but I really don't see any real improvement in flag control.  Honestly, if the ai got good at flag control, I'd think this mod is pretty damn near complete.  And at that, I simply mean if that ai ran back and forth to recap flags, that would be huge.

Thanks again for your work on this mod, hunger!

At this point, the mod is designed for team play and works best for 2v2 and larger teams (for clarity, I've inserted this in the description above).  Since I never play 1v1, I haven't observed the AI's capping behavior under this condition.  However, I am not surprised that a single AI player would lose the flag-capping race against a human player.  The AI prioritizes portals, then whatever flags are nearest and random flags.  On most maps, I have customized it to go after certain flags more often.  I think that making the AI more competitive regarding flag capturing in 1v1 is going to take some significant work, if it's indeed possible. On Nightmare difficulty, the AI has a lot of competing priorities to think about that will constantly distract it from flag capping.  For example, since it has more gold, it wants to shop a lot.  It starts hanging around the artifact shop sooner than a human player would, and its desire for artifacts and some of the other high-end gear gets pretty high once it has the gold.  Also, citadel upgrades are very high priorities, starting with FS1 and Currency.  Moreover, the AI knows that it needs to farm lanes in order to get experience, and this can take precedent over flag capping.  When all of these priorities are divided between a few different AI teammates, the AI does much better.  There may be a way to balance all of this for 1v1, but I'm not sure.   

Thanks for the feedback, guys.   My hope is that this mod becomes a community mod, where different people continue to work on improving it. 

OK, I tried to reproduce Pacov's experience with Rook vs. Regulus in Nightmare 1v1 on Cataract.  Regulus did exactly what Pacov experienced: almost no flag capturing and preoccupation with shooting towers.  Very strange.  So, I decided to try Rook against other demigods under same conditions (Nightmare, 1v1, Cataract).  I played against Sedna, UB, and Erebus.  All three went after the flags.  By the end of the game, Sedna got 18 flags, Erebus had 21, and UB had 26.  The problem seems restricted to Regulus, though I haven't tried all of the demigods in 1v1.  

Sedna, UB, and Erebus had strong early games when it came to flag capping and were quick to recapture any lost portals or other flags near their stronghold.  I still won the games handily, but not because the AI demigods would not capture flags.  I think the problem lay in the fact that by mid-game, the AI is concerned about other priorities, especially shopping.  Thus, while it would capture flags, it would then immediately go back to the citadel to snatch up gear.  It all fell apart for them once my Rook had destroyed most of their towers and then sent priests and angels which tied down the AI demigod in front of its citadel. 

I really think that there should probably be a separate 1v1 version of this mod.  While I don't really play 1v1, I was struck by how different it felt compared to, say, a 3v3 or even a 2v2.  There were very few confrontations with the AI demigod since we were both preoccupied with capping flags, shopping, etc.  Thus, in the few games that I played, there were no deaths or kills.  The games had a cat-and-mouse quality about them, where you are primarily playing tag with flags.  In team games (esp. 3v3, 4v4, or 5v5), the AI teams play much better since they can multitask much easier.  I suppose the AI could be made to do this better in a 1v1, but the cost would be decreasing its desire to shop--which, of course, would mean that it would have to give up the fighting advantage that it gets for constantly upgrading to the best gear. 

If someone wants to try this out, they would simply need to go into the AIGlobals.lua (found in hook/lua/sim/AI) and lower the PriorityFunction variable for artifacts (many are presently set for 100) and for citadel upgrades.  You would also need to go into the AIblueprint.lua of each demigod (found in hook/units/heroes) and lower the priority variable for gear associated with each demigod custom build.  After doing this, the AI demigods should be less interested in citadel upgrades and item purchased and more concerned about flags. 

For now, I'll leave the .23 version of the mod alone until people have a chance to do more play testing and to report results.  Beyond the issue with 1v1, it would be good, as Orion said, for us to compile a list of issues that need to be addressed in future versions of the mod.  In your reporting, be sure to mention the conditions of your game (that is, number of demigods, map, difficulty level, etc.).

I'm not sure that priority alone is what is causing shopping frequency.  Unless everything in the AI's shopping priorities (both upgrades and items) returns 0, or it doesn't have enough money for those that return > 0, it'll still go and shop for that one item/upgrade when it decides to shop.  What I haven't figured out by looking through the AI, is when / how frequently it decides to check priorities, and if that can be controlled (modified by, say, how much gold it's accrued or what other goal priorities currently are).

The best example of this is the perpetual desire to teleport / run back after capturing flags.  From what I can tell, your modifications lessened this, but it still happens.  It's as though they're re-evaluating their goals after completing one, and somehow shopping ends up really high, especially for nightmare AIs (implying that it must already have something to do with gold, even if it's just that they are more likely to afford things).

This should be possible.  The tournament does allow for UI mods so that you can see your allies' gear.  I haven't figured it out yet, but the tournament AI should be moddable as well.

Hi, I'd like to send you a very big thank you note for continuing the work on this mod. I've been playing Demigod over at a friend's house and I must say that this work is truly great for making all our Demigod games truly compelling.

As far as performance goes (and we've not tested it with 1vs1) the only quirky thing we've found is that sometimes enemy Demigods go through a *lot* of flak just to try and cap a portal. Like Nemado said, most times they're good enough not to get killed by the towers nearby... any other intervention, however, and they're toasted.

Thanks again for all the work!

Yeah, but for the most part, CIT upgrades are not so complicated. They can't be purchased before the war rank is reached, and they are rarely more valuable then buying the standard items.

I'm not saying you didn't so an awesome job, but is seems like misplaced effort.

It would have been better if you directed this effort towards item buying decisions instead.

eg: Priority of Narmoth's ring is 0 until you have Unbreakable / Nimoth / Vlemish, then it becomes a high priority.

Often the Nightmare AI has the gold to buy expensive items that do not give as good a return on investment as the standard 1500-1750 gold items players will always buy first... but it will purchase them first anyways because the priority is higher.

This leads to some funky orders of purchasing by the AI, because for sure the item is better, but should be an upgrade once you've filled out the efficient items, not a buy this first item.

I'm just projecting... I have 30+ new skill builds for CrazyTown, and items that compliment said builds. Also debugging the AI files is a pain, it simply tells you the AI file is invalid.

It is about a week's worth more work then I anticipated. But hey, the good thing is when I release it, there will be a template in there that includes many items. Anyone who wants to tweak the items the AI can choose simply has to copy paste it into the item section of the builds.

Idea for selling:  An optional list for which items to sell when buying each item, that the sell weighting function checks (if present).  This would create some more logical upgrade paths without too much work (though still necessitating a return to a global item weight table), and would allow the AI to avoid selling items that may complement the one it's trying to purchase.

Rook could be ALOT better by pooping out more towers.

He should be doing it non-stop when
1) He is capping a flag
2) He is fighting creeps
3) He is getting shot up by a tower

Been looking at this too, it is hard to follow, since the tournament mode is basically calling functions to populate the tournament's data. Somewhere in there, when it is populating the AIs, there must be a way to also make it use the builds.

I also saw a setting that sets the tournament to 8 demigods and a whole crap load of scripts to pick 8 of the 10 demigods, wonder if we can make it 10? That might be fun.

Also, another thing I want is to set the team AIs to the same difficulty level.

Having nightmare allies fighting nightmare enemies might make a battle easier to win, but it will make it much harder to win the tournament. As it is now, the AI loses the favor point battle because one battle it is tearing up the map and scoring high , and the next battle it is a docile pussycat on your team and gains no score.

Well, I traced the single player launch process up to LaunchSession(SetupSinglePlayerSession(sessionInfo)), which is an engine command that does who-knows-what.

Do you have to do anything special to use UI mods in single player?  If not, there should be code somewhere that differentiates between UI and sim mods for a single player session, and it should just be a matter of finding out what happens after LaunchSession, lua-wise.  This is a part of the engine I don't have much experience with - I know some lobby code, and plenty of in-session stuff, but not what order the whole session init process happens, or how it's different in SP vs MP.

Edit: LO, when I see a tower Rook in .23, he's always, always raising towers.  During an average mid-game fight, he pops at least 2-3 of them, and usually chains them out properly when advancing on a creep wave.  Are you sure you're not doing something that's disrupting this?