I know that the AI is not finished now, but i have a few questions about the future. - Land Behavior For Example SupCom, even if you have played a Island Map, the AI build Land Units and send them to the coast, where they clumped together. Will the future PA AI be able to detect something like "if surrounded by water, do not build Land Units" or use Transporters and Unit Cannons more often in these situations? - Player Detection I have never liked the fact, that most AIs know where the player is all the time. Also it seems like it is there with our current AI. So, would it be possible in the future to let the AI Scout for our base first, and if it finds us, it will attack? Something like random moving around and gather Informations of possible targets, like a player would do. - Pathfinding Most AIs will run straight into playermade traps, like turrets, ambushes etc. Lets say the AI attacks our base, will they try to avoid turrets or find the weakest spot to attack? or use longrange units (next point) to destroy them? (based on AI difficulty) - Adapting If a player builds a huge fortress, huge armys of land units etc, will it be possible for the AI to adapt to this? like countering a fortress with long-range units instead of wasting wave after wave of small tanks, use Bombers and Gunships to destroy the ground-army instead of a wasting a small land-strikeforce, build more Land/Air defense based on the players previous attacks? (based on AI difficulty)
First of all, I'm not a dev, but Sorian is a brilliant person and he has access to the game's source code now, so I think we can predict some of these. -Land Behaviour: Easily doable with a simple pathfinding check, once the AI finds its enemies. -Player Detection and Pathfinding is a technical behavior, better Sorian answers these. -Adapting: SupCom's AI did this in campaigns, to a rudimentary degree. If you built mostly air it would counter by spamming Swift Winds on certain levels. However a large table of counters to certain scenarios would be necessary.
Land behavior: I will be working with Elijah for these kinds of features. A simple: "can I path by land from point A to B" check will work wonders. Player detection: LMAO! The AI has no intelligence right now whatsoever. The AI sends groups of bombers out on random patrols. It has no concept of enemy bases and intel at all at the moment. In the future, I plan to have players be able to set the option of whether the AI gets full map vision or not. Right now that option would not make a difference. Pathfinding: That is what the influence map and neural networks are going to be for. Adapting: I love AIs that change what they are doing based on what the player does. This is definitely on my list of things to do.
Very good to know, sounds like a funny AI plan What exactly is an Influence Map and Neural Network for?
Here is a good article on influence maps: http://aigamedev.com/open/tutorial/infl ... mechanics/ Basically, it gives the AI a representation of the state of the game. Who controls what area, where the dangers are, where to attack. The influence map for PA will be different that the ones in the article, but the goal is the same. My GDC talk on neural networks: http://www.gdcvault.com/play/1015667/Of ... n-Path-Non (skip forward to around 17 minutes) Neural networks will be used to help the AI determine where to attack and when to run away.
How big are the patrol paths? If they cover the whole planet, then they will ALWAYS find the player anyways :lol:
Gotta say, that stuff on neural networks: Awesome. Not to sound like a fanboi or anything, but I swear everytime something comes around about the AI I get a little more excited. Also loved how you could easily tweek the ai by adjusting input values and such, and not by hamstringing the actual neural process.
There is a platoon task that picks a random spot on the planet and patrols to that point. Once the platoon reaches that patrol point it picks another spot on the planet and patrols to that point. So, the bombers will eventually encounter something.
Sorian your neural network talk is really cool. The ACU bug was gold. I must say I didn't really get around to putting the SupCom2's AI through it's paces, perhaps I'll pay it another visit. Unfortunately the whole economy changes thing was upsetting so I never managed to get hooked like I'd hoped. As far as AI goes though, was the SupCom2 AI limited at all either due to performance or dev-time constraints that you plan to have another crack at with PA?
that's really a very interesting talk, i have never thought that the AI was that complex. Now I'm really exited about the PA AI
I was not the primary dev on the SupCom 2 AI, I only took it over after the game was released. So, while I was able to make some major fixes and improvements, I was only able to do so much.
After training the neural networks, as everyone will share the same AI database and server, would the neural network(s) continue to learn?
