Emergent flocking boid-based crowd behavior through generalization of the system rules in a 3D Reinforcement learning environment with predator satiation and foraging
Author
Ivanov, Georgi Nikolaev
Term
4. term
Education
Publication year
2020
Submitted on
2020-06-04
Pages
65
Abstract
Formålet var at skabe fremkommende (emergent) adfærd i en virtuel crowd baseret på boids, en klassisk model hvor simple regler får grupper til at bevæge sig som en flok. Projektet generaliserede systemreglerne og indførte både rovdyrtrussel og fødesøgning, så de virtuelle agenter skulle balancere mellem at undgå fare og opsøge ressourcer. Der blev gennemgået litteratur om boids, byttedyr/rovdyr-adfærd i naturen, emergens og selvorganiserende systemer. Et 3D-forstærkningslæringsmiljø blev opbygget i Unity ML-Agents, og to maskinlæringsbaserede crowdmodeller blev trænet og sammenlignet med et boids-baseret kontrolmiljø. Resultaterne viste, at én af ML-modellerne udviklede en stabil, emergent crowdadfærd ud fra de generaliserede regler. På trods af hardwarebegrænsninger opfyldte resultaterne projektets kriterier og gav nye indsigter i, hvordan simple regler kan føre til komplekse gruppebevægelser.
The project aimed to produce emergent crowd behavior using a boids-based approach, where simple rules lead to flock-like motion. The system generalized these rules and added both a predatory threat and foraging, so virtual agents had to balance avoiding danger with seeking resources. The work reviewed research on boids, predator–prey dynamics, emergent behavior, and self-organizing systems. A 3D reinforcement learning environment was built with Unity ML-Agents, and two machine learning crowd models were trained and compared against a boids control environment. Results showed that one model produced stable emergent crowd behavior using the generalized rules. Despite hardware limitations, the outcomes met the project criteria and offered insights into how simple rules can generate complex group patterns.
[This abstract was generated with the help of AI]
Documents