💻 computer science

Genesis: An Empirical Platform for Studying Open-Ended Evolution Without Fitness Functions

The paper introduces "Genesis," an open-source platform that empirically demonstrates constraint-driven selection can sustain open-ended evolutionary dynamics without fitness functions, while also establishing that niche construction alone is insufficient for breaking complexity plateaus and suggesting a new direction for the meta-evolution of physical laws.

Original authors: Anushka Sharma

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anushka Sharma

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Escape from the Scoreboard

Imagine a video game where the only way to win is to get the highest score. You run, jump, and shoot, but every move is judged against a single number: your points. If you don't get more points, you don't get to play the next level. This is how most computer programs that try to mimic evolution work today. They are obsessed with a "fitness function"—a strict scoreboard that tells the computer exactly what to build and how to improve. But if you look at real life, nature doesn't have a scoreboard. A bird doesn't evolve a better wing because a judge gave it a higher score; it evolves because it needs to survive the wind, find food, and avoid being eaten. There is no single "best" way to be alive, just a million different ways to keep going.

Scientists have been trying to build computer programs that can do what nature does: create endless new and surprising things without a scoreboard telling them what to do. This is called "open-ended evolution." The big question is: if you take away the score, does the computer just stop trying, or does it find a new way to keep growing? This is the mystery that the paper "Genesis" sets out to solve. It asks: Can we build a digital world where life evolves just by following the rules of physics and competition, without anyone telling it what to achieve?

The Genesis Experiment: A World Without a Scoreboard

The researchers built a digital playground called Genesis. Think of it as a virtual petri dish filled with tiny, digital creatures. In almost every other experiment like this, the computer acts like a strict teacher, grading every creature and only letting the "smartest" ones reproduce. In Genesis, the teacher is fired. There is no score, no goal, and no "good" or "bad" answers. Instead, the creatures have to survive based on three simple, physical rules:

The Physics Gatekeeper: Imagine a bouncer at a club who only cares about your size and energy. If a creature's body (its "genome") gets too bloated or uses too much energy, the bouncer kicks it out. The creatures have to be efficient to stay alive.
The Adaptive Regulator: This is like a thermostat for the whole population. If too many creatures are dying, the rules loosen up a bit to let more survive. If too many are living, the rules get stricter. It keeps the population healthy without picking favorites.
The Immune System: This is a librarian that keeps track of unique creatures. If one type of creature becomes too common and boring, the librarian throws some of them out and brings back older, unique ones from the archives to keep things interesting.

The team ran these digital creatures through over one million generations to see what would happen.

The Results: Survival is Possible, but Growth is Hard

The first big surprise was that life did keep going. In 7 out of 12 different runs, the creatures continued to evolve and change even after the scoreboard was completely removed. This is a huge deal because it proves that you don't need a "goal" to keep evolution moving; you just need physical limits and competition. The creatures were fighting to stay efficient and diverse, and that was enough to keep the engine running.

However, the story hits a wall. While the creatures kept changing, they didn't seem to get smarter or more complex in a lasting way. In the first set of experiments, the researchers measured the complexity of the creatures' bodies and found that it hit a structural ceiling, plateauing at a level between 140 and 155. It was like the creatures were running on a treadmill: they were moving, but they weren't going anywhere new.

To figure out why they hit this wall, the researchers tried a second experiment. They gave the creatures the ability to change their environment, like a bee building a hive or a beaver building a dam. They thought, "Maybe if the creatures can build their own world, they will keep getting more complex." They ran 20 experiments (10 real, 10 fake "sham" controls where the creatures tried to build but nothing actually changed). The result was a hard "no." In this second experiment, the complexity metric remained at 0.00 in both the real and sham conditions. Even when the creatures could change their world, their complexity stayed flat. The environment changed, but the creatures didn't seem to notice or adapt to it in a way that made them more complex.

The Next Step: Evolving the Rules of Physics

The researchers realized that maybe the problem wasn't the creatures, but the rules of the world they lived in. If the laws of physics (like how energy works or how things move) are fixed and boring, maybe the creatures can't evolve past a certain point.

So, they are now working on a new version, Genesis V4, where the laws of physics themselves can evolve. Imagine a world where the rules of gravity or how fast things decay can change and improve over time, just like the creatures do. They are preparing a massive experiment with 100 different worlds to see if letting the "rules of the game" evolve is the secret to unlocking truly endless, open-ended evolution.

What This Means for Us

The paper doesn't claim to have solved the mystery of life or created a perfect artificial intelligence. Instead, it offers a new map. It shows us that we can build systems that evolve without a scoreboard, but it also warns us that simply removing the score isn't enough to create infinite complexity. We might need to let the very laws of the universe change along with the creatures. It's a playful, rigorous look at how life might work if we stop trying to grade it and start letting it play by its own rules.

Technical Summary: Genesis – An Empirical Platform for Studying Open-Ended Evolution Without Fitness Functions

1. Problem Statement

Evolutionary Computation (EC) has historically relied on a scalar fitness function to score solutions and guide selection. However, biological evolution sustains complex, open-ended dynamics (OEE) without such a function. Prior to this work, there was no empirical characterization of what happens when fitness is completely removed from an artificial system, nor was there a clear understanding of which architectural components are necessary to sustain evolutionary dynamics in the absence of designer-specified objectives. This paper addresses the gap by asking: what does an artificial system need to sustain evolutionary activity after the complete removal of fitness?

2. Methodology: The Genesis Architecture

The paper introduces Genesis, an open-source platform designed to test hypotheses regarding fitness-free evolution. Genesis replaces traditional fitness-based selection with three interacting mechanisms that regulate viability and diversity rather than performance:

Constraint-Driven Selection (CDS): Agents possess variable-length, codon-based genomes. Survival is determined solely by satisfying immutable physical constraints. A metabolic cost function, $M(G) = \alpha|G|^\beta + \sum k\gamma_k f_k(G)$ , penalizes genome bloat. A "Physics Gatekeeper" enforces hard resource limits, and survivors are selected via Pareto relational dominance across internal axes (energy efficiency, resource acquisition, interaction stability).
Constraint-Adaptive Regulation Principle (CARP): This mechanism dynamically adjusts constraint intensity ( $\lambda$ ) to maintain population viability ( $V(t)$ ) within a target corridor (70–90%). The update rule is $\lambda(t+1) = \lambda(t) + \eta(V_{target} - V(t))$ . Crucially, CARP regulates capacity without encoding preferences for specific behaviors.
Artificial Immune System (AIS): To prevent stagnation, the AIS maintains a novelty-based archive of representative genotypes. It prunes dominant lineages and reintroduces archived genotypes when population novelty falls below a minimum threshold ( $N_{min}$ ).

Instrumentation (PNCT):
The platform utilizes the Population-level Non-fitness Characterisation Toolkit (PNCT) to quantify evolutionary activity without fitness metrics:

Genetic Activity Coefficient (GAC): The fraction of genome edits persisting beyond a 500-generation horizon. A run is considered "sustained" if the median GAC $\ge 0.1$ .
Expressed Phenotype Complexity (EPC): The LZW-based compression complexity of expressed phenotypes. A positive EPC slope is the stricter criterion for open-ended growth.
Novelty Network Density (NND): Measures local density in phenotype feature space to ensure diversity.

3. Experimental Results

The paper reports on three distinct experiments totaling over one million evolutionary generations:

Experiment 1: Fitness Removal (V2)

Setup: 12 independent runs of 10,000 generations with progressive elimination of external fitness ( $p_m=0.01, N=512$ ).
Findings: 7 out of 12 runs (58.3%) sustained evolutionary activity after complete fitness removal. This significantly outperformed four baselines (Random Search, Fixed Constraints, Novelty Search, and MAP-Elites) with $p<0.01$ and Cohen's $d=1.47$ .
Ablation: Removing either CARP or AIS raised failure rates from 41.7% to over 90%, confirming both are critical.
Limitation: Despite sustained activity, EPC plateaued at 140–155 in all successful runs, indicating a structural ceiling where evolution continues but does not necessarily become open-ended.

Experiment 2: Sham-Controlled Niche Construction (V3)

Hypothesis: Organism-writable environments (niche construction) would break the EPC ceiling.
Setup: 20 runs (10 real, 10 sham) of 50,000 generations. Agents could secrete a persistent chemical ( $S$ ) into a 3×3 neighborhood. The "sham" control executed the action and deducted cost but did not update the chemical field.
Findings: While the secretion physics functioned correctly ( $S_{mean}=0.48$ in real runs), EPC remained zero in both conditions, and LZ76 complexity was identical (0.068).
Conclusion: Environment modification without frequency-dependent selection is insufficient to drive open-ended growth; agents' internal policies did not respond to the chemical field.

Experiment 3: Speciation-Protected Niche Construction (V4 Stage 3)

Hypothesis: Novel structures require protection (via speciation) long enough to accumulate.
Setup: Added CPPN genomes and NEAT-style speciation to the architecture.
Findings: Preliminary validation runs (two seeds, 500 generations) showed divergence in genome size and species count between real and sham conditions, unlike in V3.
Status: These results are consistent with the hypothesis but are currently insufficient to establish it; they serve as directional scaffolding for medium-scale replication currently underway.

4. Key Contributions

Empirical Boundaries for Fitness-Free Evolution: Demonstrates that constraint-driven selection can sustain evolutionary activity without fitness, though not automatically open-ended.
Sham-Controlled Protocol: Provides a rigorous method for isolating niche construction from its energetic costs, revealing that niche construction alone does not break complexity plateaus.
PNCT Metrics: Introduces a suite of falsifiable metrics (GAC, EPC, NND) to characterize evolution without scalar fitness.
Hypothesis on Substrate Evolvability: Suggests that for true OEE, the laws governing the system (the physics) may need to be evolvable, not just the population dynamics.

5. Significance and Outlook

The paper claims to establish empirical boundaries for fitness-free evolution. The observation of an EPC plateau in fixed-physics environments (V2–V3) suggests that fixed physics may be structurally unable to support OEE. Consequently, the authors propose a new direction: meta-evolution of physics, where the laws governing the system (e.g., diffusion, decay, energy injection) are themselves evolved.

The work contributes to the field of EC by providing a portable toolkit (PNCT), a reusable experimental protocol (sham-control), and a hypothesis that substrate evolvability is a necessary condition for open-endedness. The authors note that a 100-world experiment to test meta-evolution is currently in preparation.