Do low-discrepancy scenarios make a twin evaluate a policy more accurately than iid domain randomisation?
research/free-qmcdr
Do low-discrepancy scenarios beat iid domain randomisation for evaluating a policy in a twin? Latin hypercubes only cut MSE 1.6–12× (rate stays n^-0.5); shifted Halton changes the rate for smooth returns (149× lower MSE at d=2) and keeps 4–24× on a success-rate jump, shrinking with d; the R_d lattice is erratic at d=5; 95% intervals from 8 randomisations cover 94–97%. Stylised.
What it found
Averaging a policy’s return over n randomised scenarios, compared across iid, Latin hypercube, randomly shifted Halton and shifted Rd lattice samplers, against exact means. Latin hypercubes keep the n−0.5 rate (MSE cut 1.6–12×). Shifted Halton improves the rate for smooth returns (slope −0.92 at d = 2, MSE 149× lower at n = 1024) and keeps a 4–24× gain on a discontinuous success-rate return, shrinking as the number of randomised parameters grows. The Rd lattice is erratic (no better than iid at d = 5). 95% intervals from 8 independent randomisations cover 94–97%. Stylised: analytic returns on a uniform cube, no simulator.
A figure from its own output
Paper, code and results
More in Digital twins & sim-to-real
- Why does the Rd low-discrepancy lattice lose its advantage at d = 5 when evaluating a twin policy?Why does the
R_dlow-discrepancy lattice lose its edge at d=5 when scoring a twin policy? Not a seed fluke (exponent −0.45 ± 0.01 over six seeds) and not a smooth dimension effect (a dip between d=4 and d=6); a√pgenerator gets −0.66 at d=5. A simple resonance score does not explain it. Stylised. - How long does a float32 twin of a chaotic system shadow the float64 one, and do its statistics survive?How long does a float32 twin of a chaotic system shadow the float64 one, and do its statistics survive? Lorenz and Rössler: float32 departs from float64 after a median 12.3 and 164 time units, like a float64 start error of about 5×10⁻⁶ (100× the float32 roundoff), but the Lyapunov exponent and time-averaged statistics agree within sampling error (stylised).
- Does a float32 simulation of a bouncing, chaotic system lose its statistics? Sinai billiard against float64Does a float32 simulation of a bouncing, chaotic system lose its statistics? Sinai billiard: float32 leaves the float64 trajectory after a median 21 collisions, like a float64 start error of about one unit roundoff, and changes the event order one collision later, but the mean free path and circle-hit share agree with float64 and with Cauchy's exact value (stylised).
- How many research READMEs admit their citations are unchecked, and which works need checking first?Audit of how many research READMEs flag a citation as unchecked: 94 of 550 (17.1%), with Dawid and Skene (1979), Zhang et al. (2019), Horowitz (2014) and Horvitz (1999) the most repeated; counts admissions only, nothing verified.
- How much does noisy scoring cost the secretary problem, and should the 1/e cutoff move?How much noisy scoring costs the secretary problem: with n = 50 the best cutoff stays at 16–17 (19 noiseless) while P(best) falls from 0.373 to 0.145 at noise 1; cutoff rules only, simulation.
- How many simulator rollouts does a sequential test save when certifying a failure rate?How many simulator rollouts a sequential test saves when certifying a failure rate: exact dynamic programming shows Wald's SPRT uses about half the rollouts of the best fixed test at either hypothesis (79/113 vs 181 for 1% vs 5%), and 6–17% fewer at its worst point in between.
All 164 Digital twins & sim-to-real projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.