The Metabolic Price of Thought: Co-evolving Brain Microarchitecture and Cooperative Survival in a Multi-Agent Reinforcement-Learning Ecology

This record accompanies a study in which the cost of cognition is made physical: a multi-agent reinforcement-learning ecology in which every agent carries an independent, recurrent Proximal Policy Optimization (PPO) brain whose microarchitecture—working-memory width, arithmetic precision, and the number of parallel deliberation lanes—is encoded in a heritable genome and evolves under natural selection. The energetic cost of each decision is not a free parameter but is computed from the agent's actual multiply–accumulate count priced at silicon energy figures, so that metabolic burden scales, as in real datapaths, with the square of memory, the square of precision, and linearly with parallelism. Learned weights are inherited Lamarckianly under Fisher-information (elastic-weight) protection and refined by cultural distillation from a collective experience buffer. Survival is a cooperative problem: a shared food store, an emergent division of labour into four genetically biased roles (forager, hunter, medic, builder), disease and healing, cooperative hunting that turns predators into food, and constructed villages. Two ecologically distinct worlds—a benign Arcadia and a hostile Ferrox—are seeded from an identical founding population. Three mechanisms emerge: (i) without selective pressure, cooperation erodes and the benign world repeatedly collapses through a tragedy of the commons, starving amid an environment that is 98–100% full of ungathered food; (ii) predation sustains the organisation that comfort dissolves, so the hostile world persists and can outgrow the benign one; (iii) evolved brain size is a multi-attractor outcome set by the survival niche (foraging societies minimise the brain, predator-hunting societies enlarge it), while under a fixed energy budget parallel deliberation is consistently pruned by selection—a signature absent where selection does not operate. The study is exploratory and offered as a set of mechanistic case studies and hypotheses, including a "cultural-offloading" reading in which a functioning society substitutes for individual brain size. The deposit includes the manuscript (Springer Nature format), the headless simulation engine and web interface, the analysis scripts, and the figures reproducing the reported trajectories.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-06-27
DOI
https://doi.org/10.5281/zenodo.20966178
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

The Metabolic Price of Thought: Co-evolving Brain Microarchitecture and Cooperative Survival in a Multi-Agent Reinforcement-Learning Ecology

Marco Galli
Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
preprint

The Metabolic Price of Thought: Co-evolving Brain Microarchitecture and Cooperative Survival in a Multi-Agent Reinforcement-Learning Ecology

Marco Galli
preprint en

Abstract

This record accompanies a study in which the cost of cognition is made physical: a multi-agent reinforcement-learning ecology in which every agent carries an independent, recurrent Proximal Policy Optimization (PPO) brain whose microarchitecture—working-memory width, arithmetic precision, and the number of parallel deliberation lanes—is encoded in a heritable genome and evolves under natural selection. The energetic cost of each decision is not a free parameter but is computed from the agent's actual multiply–accumulate count priced at silicon energy figures, so that metabolic burden scales, as in real datapaths, with the square of memory, the square of precision, and linearly with parallelism. Learned weights are inherited Lamarckianly under Fisher-information (elastic-weight) protection and refined by cultural distillation from a collective experience buffer. Survival is a cooperative problem: a shared food store, an emergent division of labour into four genetically biased roles (forager, hunter, medic, builder), disease and healing, cooperative hunting that turns predators into food, and constructed villages. Two ecologically distinct worlds—a benign Arcadia and a hostile Ferrox—are seeded from an identical founding population. Three mechanisms emerge: (i) without selective pressure, cooperation erodes and the benign world repeatedly collapses through a tragedy of the commons, starving amid an environment that is 98–100% full of ungathered food; (ii) predation sustains the organisation that comfort dissolves, so the hostile world persists and can outgrow the benign one; (iii) evolved brain size is a multi-attractor outcome set by the survival niche (foraging societies minimise the brain, predator-hunting societies enlarge it), while under a fixed energy budget parallel deliberation is consistently pruned by selection—a signature absent where selection does not operate. The study is exploratory and offered as a set of mechanistic case studies and hypotheses, including a "cultural-offloading" reading in which a functioning society substitutes for individual brain size. The deposit includes the manuscript (Springer Nature format), the headless simulation engine and web interface, the analysis scripts, and the figures reproducing the reported trajectories.

Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

The Metabolic Price of Thought: Co-evolving Brain Microarchitecture and Cooperative Survival in a Multi-Agent Reinforcement-Learning Ecology — Marco Galli · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS