What do you tweak to make the GA evolve an interesting, fun opponent? How strong are your evolved opponents at this time? (I'm not sure evolving them sufficiently is going to be computationally cheap.)
I have two fairly separate "games" within one game. They are economy and warfare. For example, for the economy, we determine first of all which of the several mid-term and long-term goals are post profitable to reach an overall victory or as close as possible. This can be for example maximising research or culture or income, etc. We then simulate turns for the mid and long term if we were primarily focusing on one of these goals. For example for research, we only build research improvements, fund new techs, etc, and forget about the economy. This results in an overreaction, and thus we see what city would suffer the most or profit the most by focusing on something else. The coefficients are thus how many cities can afford to focus on different aspects of the economy. We simulate turns in advance (assumption is: warfare won't influence outcome much, and if it will it's compensated) and then we can try out the "best fit" for the number of cities that would focus on each aspect (eg: research vs income) to get best overall results. This relies on the fact that each of these items can be quantified, as each city can build something else, and they can trade on the empire-level. Then after finding out the extreme and mid level scenarios for research (example), we finetune by experimenting with policy changes and diplomatic influences etc, trying to predict opponents intentions in some way and so on. These factors compensate and create a list of probabilities and future reactions to possible scenarios. These are re-calculated every 10-20 turns, and the mid-tern and long-term policies adjusted etc.
In terms of warfare, we create virtual units assuming the enemy has placed them and then simulate possible outcomes depending on how the units are arranged (who is in front, who is in back, support units, ranged units, etc). We simulate potential front lines developing, and how to respond to various concentrations of forces and several scenarios, but not too many. The outcome is often a noisy affair, with a wild variety of probabilities, but then the risk-taking of the AI kicks in just like a human player would, and it tries it's best without getting fixated on any particular goal.
@IADaveMark
Yes, we have the ability for each mod to be "learnt" by the AI. This is when the coefficients are found, and it's a somewhat computationally intensive process but not as much as we expected. It still only gives guidelines for the overall game, such as build queues, overall long-term cost of any type of victory (some mods can loose balance, and this is a good way to check if the mod itself is balanced), etc. During gameplay, we still let the AI simulate turns in advance for the economy where the assumption is no external influence will be too great - but all of them are taken into account as probabilities (trade, diplomacy, war). Each simulated turn takes into account the pre-made coefficients, and tries things out depending on the current state of the empire and available resources. It doesn't test too many scenarios because it has the coeffieicent and "kind of" knows which way to go in order to achieve specified mid-term and long-term goals. There is still some testing done however, as to the order in which things should be built and so on, but these are all simple operations and the amount of data is kept to a minimum because of the aforementioned assumptions. It's very game-specific and I'm not sure if it has been done in exactly this way before, or that I'm explaining myself all that well for that matter - I'm kind of tired at the moment.