Abstract
Non-player character (NPC) control spans multiple algorithm classes, from scripted logic to
data-driven learning. This dissertation presents a Design Science evaluation of NPC control
across five representative approaches in a single Unity built 2D platformer: a designer authored
Behavior Tree (BT), a Genetic Algorithm (GA) planner, a Genetic Programming evolved
Behavior Tree (EBT), deep Reinforcement Learning (RL) with Proximal Policy Optimization
(PPO), and a Variational Autoencoder–initialized Reinforcement Learning agent (VAE-RL).
The research problem addressed is the lack of a controlled, cross class comparison that isolates
practical trade-offs among reliability, peak performance, and scalability in real time game
settings.
Within a unified observation/action interface and common scoring scheme, each controller is
demonstrated on three levels of increasing difficulty and evaluated over repeated runs. Primary
metrics are total game score (combining collectibles, enemy defeats, and time/progress) and
completion rate, with variability summarized via run-to-run dispersion. The Design Science
process guides artifact design, implementation, demonstration, and rigorous evaluation under
identical conditions and a matched training/compute budget for learning based methods.
Results indicate clear class dependent trade-offs. The hybrid VAE-RL obtained the highest
scores and maintained consistently high reliability across all evaluated levels. The scripted
Behavior Tree was also consistently reliable and remained close to the top in score, illustrating
the strength of handcrafted, interpretable control when content is fixed. Deep RL reached
competitive average scores but showed greater run to run variability and lower reliability on
the more demanding levels. The evolved Behavior Tree performed strongly on the easier
content and showed occasional failures on the hardest level. The genetic algorithm
underperformed on the trap focused medium level, where fixed action sequences are sensitive
to timing, and produced steady mid-range results on the hard level.
This dissertation contributes a controlled cross class benchmark for NPC control in a real time
platformer that uses a shared interface and unified metrics. It presents a practical hybrid
method, VAE initialized PPO, which provides a distributional warm start and improves both
stability and peak performance. It also distils design guidance for selecting and combining
algorithm classes to balance reliability, interpretability, and adaptability in deployed game AI.