City Research Online

Abstract

This thesis investigates functional creativity as a mechanism for generalisation in reinforcement learning. Its main proposal is that agents can adapt to novel tasks by recomposing structured representations of past experience, rather than relying on reactive policies or overfitting pattern interpolation. Inspired by cognitive models of creativity, we introduce a framework (AIGenC) that conceptualises generalisation as the creative reuse of abstract knowledge. Under this view, functional creativity is formalised as the agent’s capacity to extract, decompose, and reassemble task-relevant substructures, such as affordances, causal effects, and object relationships, into new behavioural strategies. This theoretical proposal is instantiated in a graph-based memory system that encodes episodic interactions as heterogeneous trajectories. Called Structurally Enriched Trajectory Learning and Encoding (SETLE), it captures high-level structure in an agent’s experience, linking actions, objects, and outcomes across time and embedding these trajectories using a hierarchical contrastive learning objective. Stored in long-term memory, these structured representations are later retrieved and matched by similarity, allowing agents to reuse subgraphs from prior episodes selectively. SETLE is then integrated into a reinforcement learning loop, where memory-based enrichment informs both action selection and policy optimisation. Retrieved graphs are filtered via attention or clustering and injected into the agent’s working memory, effectively conditioning decisions on structurally relevant past experiences. This enables zero-shot reuse of behavioural knowledge and promotes sample-efficient learning. The complete system is evaluated in two domains: CREATE, a continuous interaction environment with tools and physical affordances, and MiniGrid, a symbolic, discrete domain with sparse rewards and task diversity. In both settings, SETLE-enhanced agents outperformed baseline models in terms of learning speed, generalisation to new goals, and trajectory optimality, validating the practical impact of structured memory and creativity-inspired reasoning. The framework is discussed in the context of both the advances and limitations of current approaches, highlighting that generalisation remains bounded by the diversity of prior experience, and unsupervised perceptual pipelines (e.g., slot attention) often fail to transfer across domains. We conclude that integrating conceptual abstraction, episodic memory, and decision-making into a unified system is a step toward more adaptive, creative, and general artificial intelligence.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Departments:	School of Science & Technology > Department of Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads