Overdue update; AI and Terrain effects

Wow. September 5th for the previous post. Long time. Overdue for an update :)

AI was the task last mentioned. Guess what! It turns out AI is hard, just as I was saying in September. It is also interesting, but maybe even moreso frustrating. So what's the current state of it? Well, it works. It works and does some interesting things. We can specify AI parameters for actions we build in the spreadsheets, and it can use them; even those with complicated prerequisites. It can build sequences of steps needed for complex actions. It can forget all that and just decide on an immediate next step. It can be guided to some extent as well; the AI units can be given preferences for types of actions and considerations.

All fine and good you say, but what can it do?

It has a reasonable combat driver; it can control commanders reasonably in combat situations. It weighs objectives, current status and positioning, enemies' status and positioning, and allies'. It will retreat when hurt, rally when it can, charge from distance, advance to stay in proximity, hold when threatened.

It can reasonably find its way to a target location, even with blocked paths and terrain obstacles and complicated areas.

It can run a rudimentary protection objective, defending multiple points and sort of separating out the defense when there are multiple AIs.

It knows how to get unstuck in various ways; it will rally or use other available methods to gather its party and venture forth.

It works in a way that is asynchronous with the game, sort of, so the AI takes game time to execute rather than real time. It can be parameterized to think slow or think fast.

It doesn't sound like much, does it?

The immediate purpose of the AI is to control the caravan opponent in scenario #2. At this point it can control it reasonably well, but it still needs a few more lessons to acheive the goal. It needs to break down paths into pieces for piece-wise evaluation of threat. It needs a partially endurant memory that understands temporality. It needs to be able to be controlled to want to retreat/recover after certain actions (or it needs to decide to do so itself emergently). It needs to gracefully handle the player harassing it without a hard engagement. Still feels like a ways away.

Here's a 3 minute clip of the AI sort of succeeding a traversing the obstacle course. A couple bugs show up pretty obviously.


In any case, I decided to do something more frivolous before I return to Caravan again. I wrote in a system to have the terrain material types execute visual effects. I also changed how the terrain renders to use the materials system that got written somewhat recently. Every now and then you need an easy task with instant gratification.

I know it's not exactly amazing; but it adds a bit more life to the terrain. Here's a clip of some new layers thrown on top of the Caravan terrain as a test.

Next time should be sooner. :p

Skirmish Update

Being sick and travel does a number on productivity for side projects. Alas this means that comparatively little has been completed in the last month.

The Caravan level's introduction, dialogue, and scripting are written and working (although it needs polish of course). The main content of the scenario is next, and that was a daunting prospect. This is the first scenario where we need a somewhat intelligent enemy. Intelligent in that it can use skills at the right times, find new paths when paths are blocked, and react to the player. The original concept was to build this within the scripting system for the scenario.

But we knew in the future that we were going to want to build procedural scenarios. Procedural scenarios can't have custom scripts to control enemies. Thus was born the need for a enemy control system; an AI one tier higher than our current AIs.

The current AIs in the system behave very primitively. It manages movement and pathing, deciding on an action to execute, handling getting to a relative range of their target, deciding on a target (through aggro/threat mechanics), etc. These operate at an individual level; each unit runs these control structures independently. One of those controls is the formation follower behavior, which keeps them trying to work together and in formation. There's plenty of complexity there that had to get handled, but it doesn't have any concept of future planning or any kind of analysis.

A while back (maybe 2 years ago) I worked on a system to control enemy formations, consisting of a set of directives which were selected through a cost function and periodically sampled. It really, really doesn't work. It kind of worked, but then it didn't. If you're wondering why the current scenarios don't have any enemy formations, this is why. Those of you who have played know there is one scenario with enemy commanders, and that they are barely barely barely functional.

So the need for an AI that runs on top of the behaviors is obvious. That's the current project. I'm building an AI Planner that looks at the current state, derives its options from available actions, derives its goals from the objectives (or script), and tries to come up with a sequence of actions to accomplish it. Then, searches over the space of sequences to find the best one.

The low level planner took a couple weeks to get working at a level I'm happy with. I'm now working on integrating the planner into the actual game. As of right now, the planner runs and the plan can be executed, but all it knows about is naive movement. It can move to its goal but basically it just knows how to dumbly right-click and hope for the best.

But it's a start. I built an AI obstacle course. When it can find its way through I'll make a vid.

.... AI is hard!