Did anyone here play the Unreal Tournament series? In those games you had an option to have the AI difficulty change according to your performance. This is also a much bigger thing in online gaming for the past five years or so, where every online service has ELO match-making. Even Starcraft II has now introduced a system whereby the AI difficulty will change depending on your victories and losses. I'd love to see something like that in Planetary Annihilation, but even more granular. Looking at SUPCOM 2 there were a number of sliders to control AI bonuses and what I'd like to see is that depending on whether I'm beating or losing to the AI it starts enabling or disabling behaviours (such as transport drops), and once all behaviours are enabled it starts enabling and increasing its various resource bonuses, etc.
Sort of offtopic but truskill is undoubtedly the superior algorithm. In particular the "slow" variant from Halo 3's ranked matches seemed to produce particularly excellent and accurate matchups. On topic I'm not comfortable with the idea of an autoadjusting AI if said adjustments includes cheating. I'd like cheating to be a very specific option and not something that can get randomly applied because I did too well last match.
As I understand the "adaptive" AI simply picked a predetermined AI personality based on map, so on a water map an adaptive would pick the naval personality if the map told it to. It didn't actually adapt.
Considering that Sorian did the AI for Supcom2 (which was fantastic and one of the high points of the game), I think he's got a lot up his sleeve for the AI here. He probably has a lot of freedom to work with. I'm excited to see his implementation of AI, and being able to adjust it's efficiency.
It would be nice if that were an option you could enable/disable. It's why I'm so excited that he's in at the ground floor this time
I have a lot, like a LOT, of confidence that Sorian will make something amazing. I mean when he is finished coding the rest of the game all by himself SKill level adaptive AI would work best on a game by game basis, I think, and also a really good way to train noobs IMO
Technically speaking he 'fixed' the AI, from my understanding he didn't create the AI itself, just made it leaps better post launch. Mike
I'm expecting something decent but not magic. I think his philosophy about AI is a bit too loose for magic.
I think Sorian's weakness is that his designs almost require the cheating element to work. Normally in an RTS AI you get an initial decent assault and once you beat that, the game is generally won. When an AI cheats it rarely makes efficient use of those resources, it does the same things it would when not cheating, just a bit faster. Sorian's AI isn't like that, with the extra resources they tend to do lots of smart things which is what makes them so fun to play against, they don't feel as rote as your standard RTS AI. However, they have never fought well at the non-cheating settings and it doesn't look like a lot of time is put into working on the AI's early game. Now it's entirely possible that's due to the level of access he's had up until now. SUPCOM 1 he was modding, SUPCOM 2 he had limited time to do a repair job. I'll be interested to see what he does here. Frankly, he deserves a medal for designing about the only RTS AI which actually plays team games as a team, at least with its AI allies. Hopefully he'll also draw on some of the Starcraft 2 HotS (though credit to Warcraft 3 AMAI which did this first) stuff and get the AI communicating with the player in team games.
I felt the chat messages were the most useless thing about the sorian AI actually. I don't want the AI to pretend it's a human. He's certainly done a much better job than just about any other RTS. But his AI's also are no overminds.
I don't want chat in the trash talk style, but I do want pings and chat in the style of the AI communicating what strategy it's pursuing, where it's attacking, where it needs assistance, etc. In other words, I want the AI to play as much as a team with me as they do with one another. As to Overmind, it was a very specific AI designed specifically to take advantage of a unit the AI could handle in a way a human couldn't to then break the balance, specifically: stacking. If the unit stats were changed the AI would have collapsed, where as with the SUPCOM 2 AI a few re-runs of the neural net would have updated it. It was an interesting experiment, but not a particularly helpful one I think for people making RTS AIs.
If you read the article there's way more to the overmind than just unit movement. But even that part represents a more analytical approach to find the correct solution rather than just letting magical NN's handle it.
