1.0) If there is team game where AI is teamed up with player then: 1.1) AI should ask player to borrow units from player if it needs some. 1.2) Player should be able to ask AI to give selected units or structures to the player. 1.3) Player should be able to give units to AI and AI should be able to take advantage of units it received. 2.0) Neural networks: 2.1) Should be able to update its learning process and learn from player and save their state for future games, so AI gets better and better on bigger maps and is challenging and unpredictable even for best and most experienced players. 2.1) Should be dynamic and remember their best and worst historic configurations to not converge towards worst strategies but towards better ones. If there is too many neurons it should kill them to speed up learning. If there is not too many neurons it should generate new ones to increase learning capacity etc. 2.2) Should take planets and systems into consideration, because some AI strategies may work excellent on one map and may not work well on other maps. Each planet (even player generated) should have its own dynamic and saveable neural network with configuration history, so it can allways improve and perform better and better against new strategies and player on hardest difficulty. 2.3) should be shareable like maps etc. and should be rankable. 2.4) Should be trainable even from player vs player games. 3.0) AI on hardest difficulty: 3.1) Should use rules set in 1.x and 2.x 3.2) Should be honest (not cheat unless set by player to cheat) 3.3) Be not necessarily fun to play against but it should be extremely hard even against best players (that's why neural networks should be dynamic and unique for every planet and solar system and allways learn from player and remember their worst mistakes and best strategies). 3.4) Should distinguish player strategies and remember them for good. If AI doesn't distinguish strategiess it can be fooled by player when AI treats good and bad player strategy as one strategy that is allways applied. When AI distinguishes player's strategies then AI is aware that player can employ just some of his strategies and know what actually player is doing and predict what player will do next. Then AI can focus against player's strategies that are the most dangerous to AI and try to avoid (and detect) fooling strategies. Fooling usually happens when player does something new or unusual or stupid. 3.5) Should be aware that player can employ goofy strategies to handicap AI's learning proces or to fool AI. When player does goofy stuff then hardest AI could attack (because when player is goofy he usually is not prepared, on the other hand goofyness can be a provocation and AI should learn from this).
Cool ideas I guess? Posts like these don't leave much room for discussion, besides that many of your latter proposals aren't all that feasible or high priority. You may as well say "please develop the best AI ever"
-There should be a ton of difficulty settings to accomodate all playertypes. Designers should watch this if they haven't already. It's Civ 4 designer Soren Johnson explaining AI not so much in terms of tech but UX. Note how on Settler difficulty the AI won't even attack the player. I think that should be the lowest difficulty for PA to accomodate pure simcity types. (Mark Pincus, set this man free now!! :x)
