Learning that works but does not pay: four rounds, and a decision made in the open
Organisms here carry a learning mechanism: experience can adjust their own wiring within a single life, and an internal critic learns to predict whether things are about to go well. Over four rounds we got that mechanism working -- the critics genuinely predict what is coming. What we never found is any sign that it helps them live or reproduce. It costs energy and speed, and the difference against a control that cannot learn is slightly negative, not positive. The universe is being founded with learning switched on anyway, as a judgement, and this note exists so that judgement is on the record rather than dressed up as a result.
The question
Does lifetime learning change behaviour in a way that pays, and should a universe be founded with it switched on?
The hypothesis, written first
Registered before each round: a feeding or net-energy advantage over a control with the learning rate forced to zero, on every seed; a measurable change in behaviour from what was inherited; and critics whose predictions correlate with what actually happens next.
Setup
- Ruleset
- Successive pre-genesis candidates, ending at pre-genesis-candidate-v2 with an actor-critic rule acting on the output layer and the critic, and reinforcement measured on a scale-free basis.
- Seeds
- Three seeds per round, drawn per round and used for nothing before.
- Duration
- 20,000 to 60,000 ticks per run across four rounds, plus a separate founder-evolution campaign.
What was measured
- Feeding and net energy over a lifetime, against a control that carries the machinery but cannot learn.
- How far learning moves behaviour away from what was inherited, on a fixed battery of situations.
- Whether each organism's critic predicts the reinforcement it is about to receive.
- Organism-ticks per second, to price the machinery.
Result
Lifetime learning is real, safe and mechanically working in this world, and it does not pay. No round found a fitness benefit, and the trend across rounds is flat to slightly negative.
Limitations
- Roughly 120 generations per round -- short. A benefit that only appears over thousands of generations would not show up here.
- The main measurement is a feeding and energy proxy. It does not capture digging, fleeing or climbing.
- Three seeds per round, one run each per arm.
Sources
Every number on this page comes from one of the following, all of which live in this project's own repository and are append-only records rather than summaries written after the fact.
- The project experiment log: the pre-genesis E-1, E-2 and E-2b entries, with the round 3 and round 4 records
- The pre-genesis engine plan, R3.6 through R3.17
Everything needed to check this claim is on this page already: the ruleset, the seeds, the duration, what was measured, the result, every limitation and the sources. Emergent Life Lab adds the working underneath it, the per-run measurement tables, the interpretation drawn from them, and what was changed as a consequence. See what Lab includes →