Its Just Too Massive!

I don't think that APM is a good measurement since it usually includes all clicks and keypresses of the player.
matizpl have an APM about 180-200 and he might be the fastest PA player: http://pastats.com/pastats/chart?gameId=41032
My APM is about 160: http://pastats.com/pastats/chart?gameId=41008
But most of my APM are probably from camera movement as I use WASD to move the camera.
A more interesting measure would be "Effective Actions Per Minute", "Orders Per Minute" or "Commands Per Minute".
"Orders Per Minute" would measure how many actual orders you give during a game. Selecting units would not count while commanding 10 units to attack an enemy would count as 1 order and setting up a line of energy plants counts as 1 order.
I think that PA should strive to have about 10-60 Orders Per Minute be competitive.
My guess is that I give about 30-60 Orders Per Minute.

In the Starcraft community apm is always purely mouse & keyboard input. The advantage to this is that it translates directly to other games, since games typically use the same input devices.

I see your point, however it doesn't really translate as well as you think as the UI of the game determine *how many* key presses / mouse clicks are required to achieve the 'competitive' level of useful actions. I'd also say that when it comes to making well thought out useful decisions I doubt anyone in any game can operate much above 100 per minute. A bad UI simply requires you to memorise a sequence to enact the command which can be done more quickly as it doesn't require much thought (a bit like the old combo sequences for Mortal Combat on the old consoles, "up-down-left-left-A-B-A-down" and such like)...

Edit: Basically I have played Starcraft and Starcraft 2 (as well as all the Warcraft games) and I think the UI is pretty archaic (Blizzard haven't really evolved it, Warcraft 1 was all about rapid keyboard short cuts back in 93 and its still like that now).

That's a stylistic aspect in chess. Some players have a strength in memorizing positions, while others prefer going off-book into new territory. I think Kasparov is an example of the former, while the current champion, Magnus Carlsen, is an example of the latter.

So, my main issue with your post is that it makes some assumptions about what AI is.

There are some specific approaches to writing an AI for PA:
1. try to create an AI that is competitive
2. try to create one which is fun to play against for human players
3. try to create an AI which has human level learning and reasoning capabilities in the domain of playing PA

#3 is the only one which is actually about Artificial Intelligence, and in fact writing human-level intelligence is the main goal of the AI community. There is a reason that we talk about chess engines, not chess AI. We're aware that we're talking about computers that calculate stuff, there is nothing clever or intelligent about them.

Approach 1&2 can both be served just by writing some domain-specific logic. You don't need to use game state trees or neural networks or any standard AI approach, because the "AI" has no need to learn or reason or be clever. Sorian using neural networks is a practical decision that is chosen because it works well in a game that is still in development with constant rebalancing and new units and so on. Neural networks are a tool, in this case.

Let's say that you would try to write a Starcraft II AI. The simplest way to approach this is to have a human update the opening book to ensure that the computer would always be in a good position entering the midgame, at which point the computer could exploit its absurdly high APM by attacking at different locations and microing perfectly. (and the micro modules could just be written by humans without a need for neural networks) That's all there is to it, and I hope it's obvious that the only real intelligence displayed here is that by the designer of the AI.

I suppose 'computer player' is a better term than AI. I was talking about a specific method to implement a computer player- designing a human-level artificial intelligence is a bit out of the scope of discussion, I would say.

I also wouldn't consider game trees in any way real 'AI' programming, I just know that its a relatively common and straightforward method to program a computer player than wins.

Though I didn't know sorian was using neural networks, that's interesting. I know very little about them, and certainly haven't ever tried to implement one. Time to research!

You're correct, game trees is just calculation and I shouldn't have listed it where I did.

...I'm going to blame the forum software, and not my habit of clicking the reply button before deleting the previous post I was writing.

I really like the idea of a worker who was tasked to build X pre-saved structures, if alive and idle, will rebuild destroyed buildings. Great idea!

I would go even further than that however! I want a toggle button on any structure that I build that can be pressed in order to basically tell the game, if this structure is destroyed, have nearest idle worker rebuild. Now that would really help with micro management.