Modelleyen: Continual Learning and Planning via Structured Modelling of Environment Dynamics

The current machine learning paradigm relies on continuous representations and fixed neural network architectures to approximate environmental structures, leading to challenges with continual learning, internal structure design, and goal-directed behavior due to overparameterization and reliance on continuous parameter tuning. This paper introduces “Modelleyen,” an alternative learning mechanism that learns environmental structures topologically in an inherently continual manner, and a planning algorithm that utilizes Modelleyen’s output for goal-directed behavior. We demonstrate the effectiveness of Modelleyen and the planner in a simple environment, and also discuss their potential for creating human-comprehensible hierarchical models in machine learning.