What ARCO/UTMM is:
ARCO/UTMM is a neuromorphic market substrate: sensors (market signals), messengers (ledger), compute (metaheuristics such as a GA), and actuators (price/allocation). Objectives are expressed as low-dimensional programmes (weights and constraints) that then steer behaviour without manual/hand crafter rules. Price formation then remains emergent and auditable.
What this PoC set out to show
-
A learning controller (GA) can replace static, rule-based pricing under shifting demand.
-
Behaviour is objective-programmable (change weights → behaviour shifts, no code edits).
-
The controller can maintain requisite variety across differing demand patterns.
-
Telemetry makes behaviour and failures (e.g., market death) visible and auditable.
Observations:
-
Selection matters: GA with selection materially outperforms no-selection variants; without selection, performance collapses.
-
Learning is present: GA improves over generations; random search plateaus on this small 5-parameter task.
-
Robustness: GA v2 runs across simple/surges/seasonal/full patterns without catastrophic failure, often achieving higher utilisation and, in many cases, higher reward than the baseline; trade-offs (e.g., higher waits) follow the chosen objective weights.
-
Objective-driven behaviour: adjusting weights in objective.yaml or scenarios predictably shifts outcomes; the baseline is objective-aware for fair comparison.
-
Observability: side-by-side dashboard shows GA vs rules (price smoothness, queue stability, utilisation, reward). Market death is flagged; reward accumulation freezes post-death to surface stalls.
Limits and further work
-
GA vs random: on a small 5-parameter space, random can be competitive; more generations/population or a richer parameterisation should widen the GA edge.
-
Long-run stability (multi-thousand steps) and parameter interpretability across many seeds remain to be demonstrated.
-
Ablations (mutation-only vs crossover-only) and on-chain safety rails (caps, anti-oscillation guards) are still to be completed.
Relevance to AIBlock L1 fee setting
Fees are another control knob. The same pattern…objective-programmable policy, learned rather than hardcoded, with full observability can likely be applied to L1 fee determination, giving an adaptive, auditable alternative to static heuristics. This will still require onchain constraints, longer-horizon testing, and governance-tuned objectives with ZKP based proof of input/correctness.
Refs:
Adaptation in Natural and Artificial Systems | PDF
I’ve put together some slides here too :