An RTS engine with a good API is an ideal platform for an AI researcher working on multi-agent planning, particularly multi-agent motion planning, and dealing with uncertainty in task-level planning. RTS games are also popular for AI development competitions, both for coursework and for organizations; see the Starcraft AIIDE Competitions for an example of an adapted RTS engine and the Google AI Challenge for a competition that uses a purpose-built engine. The occasional questions about AI adaptability and personality also raise the idea of community-contributed AIs, which would certainly give bots a bit more flavor. Finally, cooperative/assistive AI is a huge help in reducing micro, simple examples being Spring's metal maker and auto-mex AIs, and could a way to introduce players to cutthroat multiplayer without either separating them from better players (limiting them to passive observation for learning from experts) or imposing handicaps. These last two work together particularly well in that all the little idiosyncrasies in players' workflows make micro-reducing assistive AI is a great target for user-generated content. Given all that, I'd like to make a conceptually simple request: it would be very nice if mods were able to provide AIs in a deep and flexible manner. This has had a couple of offhanded mentions deep in other threads, but since it sounds like AI development is this week I figured I'd put together a slightly more detailed argument for the idea.
Well, I'd like to say for one, that this game will likely make an excellent source for learning about AI because of the Neural network approach that I keep hearing whispers of. And that it will likely also make an excellent platform for AI modification, for the same reason, as well as the flow field pathfinding will likely make for some interesting possibilities (I must admit that, even though I started posting in this forum because of flow fields and misguided notions of the cost field that cropped up two live streams ago, the flow fields are still an impressively powerful tool for AI). So yeah, mod like crazy if you can wrap your head around that stuff. But I don't expect it to become widely accepted to have AI mods floating around as some kind of crutch for beginner players. And the reason why, I believe, is because Planetary Annihilation is taking a step towards shrinking the power of a "perfect" micromanipulated and maintained Economy in a way that I've only just started to grasp following the recent live streams, and some of the recent arguments that have been kicked up. It is beginning to seem to me like every decision Uber has made has also crafted a paradigm shift away from the classic RTS school of thought, that an economic powerhouse is a powerhouse no matter what, that if you are better than anyone else at over-arcing strategy, you could still become better by increasing your economy production. I'd like to get into this more, but I'd rather not derail a good thread about AI with my latest thought process that with PA, a lack of automation may no longer translate into an increase in micromanagement.
It's deeper than potential fields versus A* or seeing working examples. For example, they're mentioning artificial neural networks, but what if I got bored and wanted to try using restricted boltzmann machines instead of multilayer perceptrons? If there are hooks in a few critical places, you could do this without disturbing the rest of the system at all; without those hooks, it might be completely impossible. Same if I wanted to try augmenting the potential fields with some multiagent controls to prevent my armies from committing rocket-assisted fratricide, or if I wanted to build an auto-explore AI for my scouts, or if I wanted to entirely replace the neural network approach with something based on partially-observable markov processes. I think I could actually get a paper out of that last one. This is the point: instead of just giving us the ability to examine potential fields and ANNs as used in practice, solid access to the engine would let people use PA to build new stuff, for classroom projects, for research, for programming competitions, or for actual use in-game. On the micro-removal topic, they've done a great job of getting micro out of the economy, but unless they do something radical with military, micro will still be useful there. If units can block each others' lines of fire, if units have splash damage, if units can't move omnidirectionally or through each other, then an army with better micro will defeat an otherwise equal opposing army. I suppose it's possible that PA will be designed such that economic success utterly dominates military, but that could also turn military into simple spamming. Those two together mean that military micro assist AIs are still useful, things like formation-holding, fighter shield management, concentration-of-fire, multiagent auto-explore, and auto-skirmish. As for assistive AI for new players, I'm working off the assumption that the 40-person drop-in-drop-out FFAs will be handled by giving uncontrolled teams to an AI to keep continuity. Under this assumption a newbie would be completely incapable of participating in such a game unless they arrive at the very beginning, because if they arrive late they will inherit a side too big for them to handle. Consider, then, the interpretation that, instead of augmenting a newbie with an AI, we are augmenting a temporary AIs with a newbie. The motivation is that it's much faster to learn by succeeding on subproblems than it is to learn by failing by different degrees on the entire problem.
