Evolving Self-Organising Agents Without Fitness: Three Falsifiable Experiments from Constraint-Driven Selection to Developmental Encoding

This paper presents three falsifiable experiments investigating whether open-ended evolutionary activity can be sustained purely through structural and developmental constraints — without any explicit fitness function. The work is grounded in the Genesis Engine, a sham-controlled co-evolutionary platform designed to isolate and study the mechanisms that give rise to sustained novelty in artificial systems.

Key contributions:

• Formalises constraint-driven selection as a sufficient mechanism for sustained agent self-organisation
• Introduces developmental encoding as a second experimental axis, testing whether genotype-phenotype mappings affect long-run evolutionary viability
• Provides falsifiable predictions for each experiment, grounding the work in rigorous scientific methodology

Relevance to AGI Safety: By demonstrating that novelty generation does not require reward signals, this work directly informs the reward hacking problem in AGI — showing that intrinsic structural constraints can substitute for, and potentially improve upon, explicit fitness landscapes.

Venue: GECCO 2026 Companion (Workshop Paper / Extended Abstract) — accepted, to appear.

ResearchGate Preprint: View on ResearchGate

Recommended citation: Anushka Sharma. (accepted). "Evolving Self-Organising Agents Without Fitness: Three Falsifiable Experiments from Constraint-Driven Selection to Developmental Encoding." In Proceedings of the Genetic and Evolutionary Computation Conference Companion (GECCO '26 Companion). To appear.
Download Paper