Inferring Complex Dynamics through Core Knowledge

The ability to generalize knowledge across similar environments is a key aspect of human intelligence. Unlike humans, current artificial intelligence (AI) models struggle to extract structured, transferable knowledge from dynamic environments. DeepMind’s breakthrough results on Atari games have demonstrated that pure sub-symbolic approaches, such as deep reinforcement learning, lack true understanding: they optimize policies based on raw pixel inputs but fail to generalize when minimal changes are introduced to the environment. This brittleness—where even trivial modifications, such as recoloring a paddle or shifting an object’s position, require retraining from scratch—suggests that these models do not learn in the human sense but instead rely on fragile heuristics. As argued by Melanie Mitchell, such systems lack an internal "world model", a structured representation of cause-effect relationships that allows for reasoning and adaptation. This thesis presents a novel framework for capturing interpretable knowledge about game dynamics without relying on opaque neural networks. Instead of passively learning from vast amounts of data, our system actively constructs structured models of object behavior using heuristic-based reasoning grounded in core knowledge. The method does not assume pre-defined objects but begins with anonymous patches extracted from game frames, identified solely by their fundamental properties (e.g., position, shape), defined by the core knowledge. A set of heuristics then tracks these patches over time, recognizing persistent entities and their interactions. Through rule inference, the system identifies regularities in object behavior, reconstructing the underlying mechanics governing the environment. These structured representations are abstract and reusable, meaning they could, in principle, transfer across related environments without requiring retraining. Unlike deep learning models, which discard prior knowledge when faced with novel conditions, our system retains and adapts its understanding dynamically. This approach lays the groundwork for AI systems capable of interpretable, human-like reasoning about interactive environments. The model can potentially be exploited as starting point for reinforcement learning (RL) frameworks, serving as an "internal world model" for agents in Dyna-like architectures. Preliminary experiments indicate that leveraging structured knowledge in planning improves sample efficiency and robustness to environmental changes, suggesting promising directions for future work.