Demigod is an innovative hybrid between RPG and RTS games, challenging the player to control and manage a Hero-like unit to accomplish a variety of missions (e.g. Dominate, Conquest, Fortress) in a set of predefined maps. It's built mostly as a multi-player experience, but includes an offline Skirmish mode for both quick games and tournaments against the AI. Since the implementation is based on a goal-oriented action planner, I've been keen to play the game and dig into the Lua code for the last week.

I was expecting to find a textbook application of the goal-oriented action planner (GOAP) from F.E.A.R. pioneered by Jeff Orkin. On the surface that seems to be the case, as the architecture and many of the high-level concepts are similar. However, under the hood, almost everything is done differently — to the extent that it's more similar to other game AI approaches than goal-oriented planning. And indeed, due to the RPG nature of the game, the long term progress of the AI Demigods, and the dynamic yet tactical gameplay, the problem required a very different solution.

In this in-depth feature, you'll learn how the AI used for the Demigods was architected and implemented (a.k.a. Hero GOAP). You'll find out why the name "goal-based action optimization" fits better than calling it a traditional STRIPS-like planner implementation, and what was necessary to make such a system work on such a large search space with a wide range of choices available.

NOTE: This first article covers only Demigod's Hero AI — which is the lowest-level and most complex layer of the whole system. The next article will show how everything fits into a three-layer architecture with multiple planners that include a squad AI and strategic reasoning.