System Modeling: The Substrate for Neural Co-Evolution
- Source
- unconv.ai
- Author
- Carlos H. Morales
- Date
Context
Training an AI model against a simulated circuit is useful only if the simulation preserves the circuit behavior the model learns to exploit. Unconventional AI explains this problem using a double pendulum, a pendulum with a second joint whose motion is easily chaotic: adding friction or changing numerical accuracy can alter where predictable motion becomes chaotic. A model trained on an idealized system could therefore rely on behavior absent from the real hardware. The modeling article uses this example to explain its simulation requirements.
The engineering challenge is to retain the physical details that affect computation while keeping simulations fast enough for repeated training. The company's stated aim is an abstraction simple enough to run quickly yet accurate enough that results from training in simulation correlate with its physical chips.
- The piece uses a double pendulum as a proxy for near-chaotic hardware: an ideal frictionless simulation and a friction-damped one started from identical conditions swing in sync briefly before wildly diverging at the 'onset of chaos.'
- The gap between an idealized simulation and a physically realistic one is largest exactly where the dynamics are most expressive, meaning a model trained only on the ideal simulation could learn to exploit a near-chaotic region that doesn't exist in real hardware.
- Numerical solver choice matters as much as physical modeling: swapping a coarse Euler-method solver for a high-order one like DoPri shifts where the simulated chaos boundary falls, since rounding error behaves like an added physical perturbation near the edge of chaos.
- The team frames its goal as finding a 'Goldilocks' level of abstraction rather than maximum fidelity, since a perfectly accurate simulation takes hours per run while their AI scientists need to run millions of passes.
Checking sign-in…
Loading comments…