In this dev blog, we will look at the Artificial Intelligence of Casual Pro Wrestling.
Unless you're in a small match with only a couple of friends in either couch multiplayer or netplay (more on these in future blogs), it is absolutely crucial to have AI-controlled wrestlers that play along. Apart from that, it's so much fun to watch AI-only matches in wrestling games!
Naturally, CPW features AI controlled wrestlers. The main idea is that they should be able to do all available actions and create a pinch of unpredictability. In the following, we will look at some details.
Virtual Controllers and Reaction Times
Fairness is likely the most important aspect when dealing with AI players. If the AI has silly shortcuts unavailable to human players, it's usually where the fun ends. In CPW, picture the AI holding a virtual controller: in order to do anything, they have to point a direction with a virtual analog stick and press virtual buttons just as a human would have to. If humans need to hold a button, e.g., to perform a heavy strikes, so have the AI players.
Further, the AI in CPW doesn't always click instantly. As human players do, they have to react to what is going on first, which may take some time (determined at random and eventually by difficulty). As a result, what may happen is that an AI player simply starts blocking too late and you still land that punch.
Goal-Oriented Behaviour
CPW's AI wrestlers act goal-oriented. This does pertain to Goal-Oriented Action Planning (GOAP) that some of you may know per se, but a variant of it that is also used for the mobs in Minecraft, to name a pretty popular example.
The AI knows a predetermined set of actions and how to perform them (i.e., what virtual buttons to press). These actions are very fine-grained, e.g., "approach the opponent", "do a weak strike", "try counter the current move". Each goal has a priority based on the current situation, so while "try counter the current move" may be of zero importance when running the ropes, it will have a high priority when the opponent is performing a move on them. Whatever goal has the highest priority in any given moment will be pursued. To spice things up, there is always some randomness involved that cause priorities to shuffle to an extent. The following flowchart sketches how the AI ticks roughly:
Action goals are completely decoupled from navigational goals. To see what this means, "approach the opponent" knows nothing about climbing out of the ring if said opponent is outside. This is handled by the navigation behaviour, which in its own knows goals like "go to the apron" and "climb off", or simply "quickly slide out". Again, the choice here is based on current priorities.
This system is extremely flexible and ridiculously easy to maintain and extend. Subtle changes in priorities may cause a wholly different behaviour, and adding or removing goals allows to enforce behaviours helpful, e.g., for wrestlers entering the ring, which is done simply by an entrance AI. At the same time, this system does not do any complicated computations behind the scenes and is thus very performant.
Grid-Based Pathfinding
Today's free-roam 3D games are probably dominated by navigation meshes, which are basically a compressed form of navigation grids that you may know from strategy games. Navigation meshes require very little RAM and allow for very simple pathfinding, and thus they seem like the best choice for a wrestling game initially.
Now let's say somebody places a large object, a table, somewhere in the ring. AI wrestlers shouldn't try and walk right through that table to reach the opponent, but walk around them. For that to happen, the pathfinding system needs to be updated. Updating navigation meshes, however, is everything but simple. The easiest and safest way is recomputing them, but this can be extremely costly.
For this practical reason, CPW uses a good old grid instead, as shown in the following picture:
Green cells are perfectly accessible, whereas red cells employ a penalty (the table could be moved if need be, but this takes longer than just walking around it). Note how some cells beneath the wrestlers are also slightly red-ish - this is to avoid that wrestlers walk through one another for no reason, all without the need for complex (and hard to get right) algorithms like Reciprocal Velocity Obstacles or similar. Grids are very easy to update and allow for well-known pathfinding algorithms like A* to be used directly.
The downside of grids, in general, is that they require substantially more memory than navigation meshes. It turns out, however, that the grid doesn't need to be too fine-grained for a wrestling game. We're talking only a bunch of megabytes in RAM, which even on horrible hardware nowadays is negligible.
To round it up, AI has been a key endeavour in CPW's development. Naturally, it's a topic that will not end with the game's release and will remain subject to tweaks and balancing for a long time. To this end, it has been designed to be easy to maintain and extend. The virtual controllers were actually meant to make machine learning a possibility, but this is definitely something for the far future when the game is in a very ready state.
For those who have been following, thank you for reading, feel free to discuss and ask questions, and stay tuned for the next update!