Chris McCole

View Original

Refactoring The Refugees and AI in Resilience

Debut Trailer for Resilience the Sci-Fi City Builder

While working on the sci-fi, civically driven, city builder Resilience, we ran into many technical challenges that are typical when developing any city builder. However, Resilience was unique in that we weren’t simulating the buildings and city planning, as much as we were simulating each resident Murian in the community. On top of this, we developed the game to be played secondarily in the top-down view typical of a city builder, and primarily as a first person experience. The team made this design decision to give the player the experience of navigating the camp along with the refugees, as well as to ground the player in reality and treat them at the same level as any of the other inhabitants. The ultimate goal in making this design decision was to help the player build empathy with the refugees of the camp. In order to build this empathy, Resilience greatly relied on the refugees being able to take care of themselves and do more than go from point A to point B. They needed to: interact with each other and the player, vary their dialogue based on the state of the camp, the health of themselves and the health of their friends, have animations and an animation controller that reflected all of these conditions visually to the player. The residents had to feel like real beings that the player could relate to and empathize with. This is in contrast to most other city builders in which the NPCs are like tiny ants in the grand scheme of the city. They don’t have to be interactable or reflect any knowledge of the area or their surroundings. Most city builder NPCs just need to look like they are aware of these aspects from afar. Thus, you can see the purpose and unique challenges the  AI for the refugees in Resilience present.

The AI and Refactor for the Refugees in Resilience was primarily directed by Me (Chris McCole) and Tejas Shah.

The residents of Resilience, Murians, live together in homes, which form communities. A Murian first arrives at the camp via dropship and they check into the camp via the entrance lines. Once checked in, they are assigned a home, or declared homeless and put onto the home waiting list. Once in camp, they need to navigate to get themselves between their home, food, water, hygiene, and medical distribution facilities, all to make ure that they, and their roommates (if they have a home), are well taken care of. When a refugee is a member of a home, they know the stats of the other Murians they live with. A home can send a single refugee to gather food or water for the whole house. This functionality makes homes a more efficient way to run your camp. All of these demands led to the requirement of AI agents that were aware of themselves, the environment, the resources around them, and their roommates.

If we had all of these knowns, we may have been able to come up with a robust, easily extendable system from the beginning of development. However, like any large game project, this idea was built up over time, over many iterations, play tests, and prototypes. It wasn’t until well into our 9 month development time that we realized the full complexity of what we needed. And as we had been building the game, slapping in functions, behaviours, and stats wherever seemed fit, whenever we needed, we ended up with a ginormous glob of coupled classes. We had one, extremely lengthy, script to represent a refugee. This included any AI behaviour, housing info, personal stats (health, hunger, thirst, shelter, hygiene), animations, UI overlays, navigational info, etc. The script was at the point where anytime we needed to update or add functionality, it took a long time to figure out where to insert the code. Once inserted, running the code had any number of crazy side effects, as the control flow was extremely hard to predict and reason about. This led to The Great Refactorization.

Determining how to refactor the AI was not a simple task, however, it involved a few major components that we knew we wanted. We wanted a component to reason about the refugees stats, a separate animation controller, a UI overlay manager, as well as the AI component. Everything had a pretty clean break between functionality, except for the AI component, in which we were unsure what information was needed, how to get it, and how to flow between behaviours. After some deliberation, we decided to go with a behaviour tree representation of the AI, as Halo is known for. This would allow us to cleanly separate behaviours and carefully define conditional checks to modify the flow between behaviours. Best of all, if we needed a new behaviour, it was as simple as making a new class, writing the code for that behaviour in isolation, and then defining when that behaviour should trigger. All we needed was to set up the inheritance for the behaviour tree components, and structure the data that needed to be passed to the tree for it to be able to determine which behaviour to execute.

In building the behaviour tree we went with a classic implementation that you could find write ups of elsewhere. For the sake of completeness, I will briefly explain the structure of our behaviour tree. The behaviour tree is made up of many nodes, each of which derives from an abstract base class. In our game, this class was called Task, all other node classes would inherit from this. Those classes being Composite, Condition, and RefugeeAction. The purpose of these classes is to give structure to your graph, while relying on the polymorphic traits that you get from the inheritance. 

The Task class allowed us to define a context of data for the class to use in it’s execution. Task also defined Enter, Execute, and Exit abstract functions, allowing all descendant classes to implement however they need. The Composite class allowed us to hold an array of Tasks, in which we could derive a Selector class and a Sequence class. Each of which looked at the array of Tasks differently, and determined how to flow between the nodes in the behaviour tree. A Selector, for example, Executes every Task in the array and stops executing, once one of the executions returns true. This allowed us to run code for one set of Tasks, effectively selecting the behaviour we wanted. Opposite to this is the Sequence class which Executes every Task until the task returns false. This allows an entire sequence to be Executed as if it were one Task, allowing us to break down large behaviours into smaller chunks. For example, we could run the Task “Check If Refugee Is Hungry” and then “Go Get Food,” so that the refugee only gets food if they are hungry. Finally, the Condition class simply returns true or false depending on if some condition is met, such as “Refugee is hungry.” Using only these base classes we developed all of the resulting refugee behaviour.

The image below shows our initial behaviour tree conceptualization to handle the behaviours we cared about. Later we added more behaviours and conditions, however, it was easy to do so, as we would develop the behaviour as a single node and insert it into this graph wherever needed. The behaviour tree graph runs from left to right, and stops executing when a child node returns successfully. This property allows us to set a priority of behaviours and tasks by structuring them from left to right.


In this graph, the simplest example is the “IsDead” action that will run when the refugee dies. This would return successfully to the root node and the tree would stop executing. To show the tasks in order of priority, I will list them here: IsDead? -> Is Player interacting with me? ->Do I (the refugee) have any immediate needs only I can take care of (Hygiene, medicine)? -> If I am in a community, do I or any of my community members need food or water? -> If all my needs, and all of my roommates' needs are taken care of, go do something idle such as going home, or to the community lot, etc. This priority ordering makes it very clear when behaviours will be executed, and when

As a result of this refactor, we were able to more easily add functionality to the refugee, debug unexpected behaviour, and isolate components. We were also able to run AI for a couple refugees at a time, to spread out the execution across multiple frames. This was done by taking the AI update away from every refugee, and giving it to the refugee manager. The refugee manager handled sending out AI and stat updates to each refugee it was aware of. The refugee manager also told other systems in the game, such as the score manager, of important refugee updates, such as when a refugee had died. 

The development of the refugee AI was not done upfront, nor overnight. We decided to use a behaviour tree only after having much of the AI functionality developed. It was also after realizing that our development, and conceptualization of the code was being hindered by the code’s structure. Refactoring was not a simple task, breaking the components apart while maintaining feature parity took a long time, lots of research, and testing. The refactor alone may have taken about 20 - 40 hours of work across two weeks. Teaching the team how the system worked took time as well. However, the benefits of code readability and modifiability more than made up for this work.The refactor made improving and adding features to the game easier and more fun. We would recommend that anyone look at the complex systems in their game, to see if they could be refactored. Even if the work adds no new functionality, it can be worthwhile for the development process to improve the readability, modifiability, of these systems for you and/or your team.

See this form in the original post