Research & Papers

Genesis platform evolves agents without fitness functions

Can evolution happen without a designer-defined fitness function? This paper says yes.

Deep Dive

In a GECCO Companion '26 workshop paper, Anushka Sharma presents Genesis, a platform where digital agents evolve on a Gray-Scott reaction-diffusion substrate solely under physical constraints. Three falsifiable experiments explore evolutionary dynamics without any designer-specified fitness function. The first experiment shows that constraint-driven selection sustains evolutionary activity after complete fitness removal but hits a hard phenotypic complexity ceiling. The second adds agent-mediated niche construction via chemical secretion, which proves real but insufficient to break that ceiling.

The breakthrough comes in the third experiment: replacing the fixed-alphabet genome with a Compositional Pattern Producing Network (CPPN) indirect encoding, protected by NEAT-style speciation. This produces the first evidence of progressive structural complexification in a fitness-free system. Null results are treated as precise answers rather than failures. The paper also contributes reusable diagnostic tools and a sham-control protocol for evaluating any open-ended evolution system.

Key Points
  • First evidence of progressive structural complexification in a fitness-free evolutionary system using CPPN encoding and NEAT speciation.
  • Constraint-driven selection alone sustains activity but hits a complexity ceiling; niche construction via chemical secretion is real but insufficient.
  • Genesis platform uses a Gray-Scott reaction-diffusion substrate as the environment for self-organizing agents.

Why It Matters

Opens a new path for open-ended evolution without predefined objectives, relevant to AI and artificial life research.

📬 Get the top 10 AI stories daily