Learning Intentions Following Evolution

Información del proyecto

LIFE

Identificador del acuerdo de subvención: 101062294

DOI

10.3030/101062294

Proyecto cerrado

Fecha de la firma de la CE 5 Septiembre 2022

Fecha de inicio 1 Septiembre 2023

Fecha de finalización 31 Julio 2026

Financiado con arreglo a

• Marie Skłodowska-Curie Actions (MSCA)

Coste total

Sin datos

Aportación de la UE

€ 181 152,96

Coordinado por

UNIVERSIDAD POMPEU FABRA
Spain

Periodic Reporting for period 1 - LIFE (Learning Intentions Following Evolution)

Período documentado: 2023-09-01 hasta 2026-07-31

Resumen del contexto y de los objetivos generales del proyecto

Understanding why something was said, and not only what the words mean, is central to human communication. When this action was proposed, such pragmatic reasoning was a recognised frontier for artificial intelligence. Large language models have since transformed what machines can do with language, but by a route this project deliberately did not take: training very large statistical models on very large corpora, once, then deploying them frozen. The questions that motivated the action remain open. How can a system learn from little data, keep adapting after deployment, and hold its knowledge in a form a human can inspect?

Rather than designing the learning machinery by hand, the project asked whether it could be made to emerge and evolve. The starting point was Combinatory Chemistry, an artificial chemistry in which elementary program fragments drift in a simulated "soup", combine at random and, being pieces of a Turing-complete language, also compute on one another. Self-maintaining cycles of reactions arise spontaneously, and some reproduce. The plan was to make that chemistry evolve by running it at scale and letting its computations make occasional errors, then to steer it by feeding the soup when it behaves usefully, giving a learning rule in which the evolving population does its own exploring.

Everything rested on the first step, a problem open in artificial life for decades: what conditions give rise to an open-ended evolutionary process, one that keeps producing novel and more complex forms indefinitely?

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

The main work focused on exploring the conditions to unlock an evolutionary process in a software simulation. For this, the fellow wrote from scratch a new simulator in the Rust programming language that adapted existing techniques from the computational chemistry literature as well as novel ones that the fellow specifically introduced for this system. This effort made it possible to simulate two to three orders of magnitude larger systems that it was possible before. Furthermore, new reaction rules like mutations were introduced to enhance the exploration capabilities of the system.

The resulting simulations produced a rich variety of complex structures. What none carried was a genetic code, namely, the separation between a heritable description and the machinery that reads it, widely regarded as a cornerstone of evolution. This led to a working hypothesis: that the absence is not a matter of simulating longer or larger, but that Combinatory Chemistry cannot express such an encoding at all. While this hypothesis is not proven, it was plausible enough to redirect the action.

The project therefore turned to Interaction Combinators, a graph-rewriting formalism from 1997 that can express a genetic code, built a second simulator for it, and characterised theoretically a self-reproducing structure carrying such an encoding. Simulation did not yield that replicator or any other, most plausibly because of the vast combinatorial search space. The learning system that was to be tested on it both depended on first obtaining evolutionary dynamics, and so could not be realised.

In parallel, a collaboration with the statistical physicist Dr. Giampaolo Folena examined how these chemistries behave in aggregate, covering the role of "artificial enzymes", reaction cycles that reshape their own chemical environment, and the equilibrium statistics of assembly networks.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

The main outcome is the new simulator, openly licensed and available at https://github.com/germank/cc3. It is usable by anyone working on topics connected to artificial life.

A second result came out of a collaboration led by the physicist Dr. Giampaolo Folena. In chemistries of this kind the number of possible structures explodes as structures grow larger, and the work asks how do they organize at equilibrium when there is nothing but random mixing driving the system. The study identifies two co-existing regimes: one with small structures which occur in many copies and behave predictably, and another one past a size that depends on how large the system is, in which each structure is so rare that whether it appears at all is down to chance. Ranked by how often they occur, structures follow a Zipf-like pattern, familiar from the frequencies of words, which here comes from the counting alone, with nothing selecting or favouring anything. The lesson is that much of the apparent order in such a system comes for free, and any later claim that something is being selected has to be measured against it. The article is under review at SciPost Physics. Together, we also proposed a model for "artificial enzymes", presented at the ALIFE conference.

A third finding concerns the difference between complexity and heredity. Combinatory Chemistry produces elaborate structures and some that reproduce, yet none carries a description that can be copied and varied independently of what it describes. The hypothesis this suggests is that computational universality does not by itself confer the capacity to evolve, and that Interaction Combinators is the more promising substrate. It is not yet published or proven, and testing it will require further research.

The potential impact of discovering a substrate that supports heredity and can be made to evolve on its own would bear on how evolutionary units originate, and would give machine learning a system whose search strategies are not fixed in advance but grow with experience. Further research is still required to reach that milestone.

On exploitation, the host university's innovation unit has been engaged and a patentability study has been requested from the Spanish Office of Patents and Trademarks.