Evolving Self-Organising Agents Without Fitness: Three Falsifiable Experiments from Constraint-Driven Selection to Developmental Encoding
This paper introduces the "Genesis" platform to demonstrate that while physical constraints alone cannot drive progressive structural complexity in fitness-free evolution, combining agent-mediated niche construction with indirect developmental encoding (CPPNs) and speciation successfully enables open-ended evolution.
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
Imagine a world where life doesn't need a teacher to tell it what to do. In the wild, animals don't have a scoreboard telling them they are "good" at being a lion or a fish; they just have to survive the heat, the cold, and the hunger. For decades, scientists trying to build digital life on computers have struggled with this. They usually give their computer creatures a "fitness score"—a digital reward system that says, "Good job, you found food!" or "Bad job, you crashed!" But this feels a bit like cheating. The big question in this corner of science is: Can we build a system where digital creatures evolve complex, interesting behaviors just by following the rules of physics and chemistry, without anyone handing them a scorecard? This paper dives into that mystery, exploring how "constraints" (like running out of energy) can act as the only teacher a creature needs, and whether adding a little bit of "developmental magic" (like how a human baby grows from a single cell) helps them get smarter.
The researchers behind this study built a digital playground called Genesis to test these ideas. They wanted to see if they could create a system where agents (tiny digital creatures) evolve on their own, driven only by the physical laws of their world, with zero "fitness function" telling them what to win. They ran three main experiments, treating every failure not as a mistake, but as a very precise answer to a question.
First, they tried to evolve creatures using only physical rules. Imagine a game where you can't die from a "game over" screen, but you can die if your internal battery drains too fast. The scientists removed all the "good job" signals and let the creatures evolve for 10,000 generations. They found that the creatures did keep evolving and changing, which was a huge success. However, they hit a "glass ceiling." The creatures got a little more complex, but then they stopped. They couldn't get much bigger or smarter, no matter how long they played. The study showed that two specific tools—a dynamic difficulty adjuster and a "novelty keeper" that saves weird ideas—were essential to keep the game going, but even with them, the creatures couldn't break through that ceiling.
Next, they asked: "What if the creatures could change their environment?" In biology, animals often build nests or change the soil, which helps them evolve. The team let their digital creatures secrete chemicals to change the world around them. They used a clever trick called a "sham control" to test this. In one version, the creatures actually changed the world; in the other, they thought they were changing it, but the computer secretly blocked the change. The result was surprising: even when the creatures successfully changed their environment, it didn't help them break the glass ceiling. The environment changed, but the creatures' brains didn't get any more complex. This ruled out the idea that just "playing with the environment" is enough to create complex life without a fitness score.
Finally, the researchers tried a different approach: Developmental Encoding. Instead of giving the creatures a flat list of instructions (like a simple recipe), they gave them a "growth plan" (like a blueprint for building a house). They used a special network called a CPPN that could grow and change its own structure, protected by a system that kept different "species" from eating each other's lunch. This time, the results were different. The creatures started to grow much more complex structures. While the first two experiments hit a wall, this new method showed the first signs of breaking through it.
To make sure this wasn't just random digital noise, they ran a final set of tests. They found that the creatures with the complex "growth plans" were actually better at solving problems and using energy efficiently than the simple ones. They proved that the extra complexity wasn't just "bloat" (useless extra code); it was a necessary tool for the creatures to survive in a changing world.
In short, this paper suggests that while strict physical rules can keep digital life alive, they aren't enough to make it truly complex on their own. Simply changing the environment doesn't fix the problem either. However, giving the creatures a way to "grow" their own brains, rather than just hard-coding them, seems to be the key to unlocking a new level of complexity. The study concludes that to build truly open-ended, self-organizing life, we need a mix of physical constraints, a way to protect new ideas, and a developmental system that lets complexity grow naturally.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.