AI Modules
Summary:
Before I was a community manager, I studied as a programmer. For my exam I made a lot of cool little modules for AI in an artificial ecosystem. I also created a simple system to gather telemetry data about the whereabouts of the critters who lived in the simulation, so we could generate a heatmap of location hotspots over time.
Below is a list of articles I wrote about each of these component systems:
Sight and memory:
A system used to allow the AI to observe and remember other game objects in its surroundings. It comes with a simple API for handling these observations, such as verifying if an object has been seen before, returning a list of observations of a specific type of objects (eg. where did I see mushrooms?), returning the most recent observation of a specific type etc. This system is used in conjunction with the AI’s decision-making processes, to create behaviors that are influenced by what the AI has previously observed.
Weight-based randomization:
This system is used for decisionmaking, which looks at the internal stats of an animal, prioritizing those that are more critical, above those that are not. There is still a small chance that non-critical behaviors are chosen, but every option is weighed relative to one another, with the critical ones increasing exponentially. This allows for fairly randomized decisionmaking, but should generally commit to what will keep the animal alive. This same system is used when generating loot drops in RPGs, or in lootboxes.
Fleeing behavior:
When faced with one or more threats, the AI calculates the optimal direction for escape. This prototype illustrates how those calculations are made, using simple vector math. The blue cube in the video is the animal trying to escape – the red cubes are what it is escaping from. The little green cube indicates where it wants to escape to. When there are no obstacles in the way, the green cube will always be at the end of the red line. But as soon as the red line collides with an obstacle, white lines make distance measurements to find out where it can move to. Smaller obstacles require multiple samples at narrower angles, to approximate where their edges are located.
Searching behavior:
A simple system to generate a field of points in the navmesh, which the AI can use to randomly explore an area. It uses easily configurable parameters to change how the field is generated. This behavior is used when an animal decides to do something, but has no memory of the desired object type (eg. “I want to eat but have seen no mushrooms/herbivores”).
Movement system:
This system is used to dynamically update the Navmesh Agent of the AI with a new path, as smoothly as possible (without anything breaking). It keeps track of the path currently being followed, as well as how far the AI is along that path. It has a small API that receives the new path, then automatically resets the index to indicate that the animal is now at the start of the new path.
Telemetry collection and analysis:
The first article deals with the collection of positional telemetry from the animals, sending it off to an online database, which is configured to properly format and store the data:
The second article deals with the retrieval of this data and visualizing it in a separate Unity client. We opted to manually download the data instead of streaming it, as this was easier to do within our time limit:
