Bibliography

The literature Auracle's methods come from, grouped by where they appear. These are the specific results the implementation relies on, not a survey.

Preference learning

Bradley, R. A. and Terry, M. E. (1952). Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons. Biometrika 39(3–4), 324–345. → The duel likelihood, $P(A \succ B) = \sigma(u_A - u_B)$. Used in

Chu, W. and Ghahramani, Z. (2005). Preference Learning with Gaussian Processes. ICML. → The framing of preference data as observations of a latent utility. Auracle's utility is linear in a fixed feature map rather than a GP, which is a deliberate trade of flexibility for interpretability and a tractable cold start.

McCullagh, P. (1980). Regression Models for Ordinal Data. JRSS B 42(2), 109–142. → The cumulative-logit model with learned cutpoints, which is how star ratings are treated as ordinal rather than as numbers. Used in

Brochu, E., de Freitas, N. and Ghosh, A. (2007). Active Preference Learning with Discrete Choice Data. NIPS. → Preferential Bayesian optimization: the loop of latent utility + expensive human oracle + cheap surrogate that Auracle's two loops implement.

Active learning and acquisition

Houlsby, N., Huszár, F., Ghahramani, Z. and Lengyel, M. (2011). Bayesian Active Learning for Classification and Preference Learning. arXiv:1112.5745. → BALD: expected information gain about the parameters. Implemented and selectable; it ties uniform pairing on this problem at session horizon.

Yue, Y., Broder, J., Kleinberg, R. and Joachims, T. (2012). The K-armed Dueling Bandits Problem. JCSS 78(5), 1538–1556. → The dueling-bandit framing, and by extension the Thompson rule that measurably loses here because it optimizes best-arm identification rather than parameter recovery.

Calibration and scoring

Brier, G. W. (1950). Verification of Forecasts Expressed in Terms of Probability. Monthly Weather Review 78(1), 1–3. → The proper scoring rule that replaced accuracy. Why that mattered

Gneiting, T. and Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. JASA 102(477), 359–378. → What "proper" means, and why a rule that is not proper can be gamed by a model that hedges.

Dawid, A. P. (1984). Present Position and Potential Developments: Some Personal Views. Statistical Theory: The Prequential Approach. JRSS A 147(2), 278–292. → Prequential evaluation: score each forecast before seeing its outcome. This is exactly what record_duel does, and it is what makes the reliability diagram out-of-sample.

DeGroot, M. H. and Fienberg, S. E. (1983). The Comparison and Evaluation of Forecasters. The Statistician 32, 12–22. → Reliability diagrams, and the calibration/refinement decomposition that explains why the shape of the failure is more informative than the scalar.

Monte Carlo

Metropolis, N. et al. (1953). Equation of State Calculations by Fast Computing Machines. J. Chem. Phys. 21(6), 1087–1092. Hastings, W. K. (1970). Monte Carlo Sampling Methods Using Markov Chains and Their Applications. Biometrika 57(1), 97–109. → The sampler.

Green, P. J. (1995). Reversible Jump Markov Chain Monte Carlo Computation and Bayesian Model Determination. Biometrika 82(4), 711–732. → Trans-dimensional moves: what a structural proposal is, since it changes the set of sites. Handled by fugue rather than by Auracle.

Del Moral, P., Doucet, A. and Jasra, A. (2006). Sequential Monte Carlo Samplers. JRSS B 68(3), 411–436. → Tempered SMC: the designed generation mechanism, and not what currently ships.

Kong, A., Liu, J. S. and Wong, W. H. (1994). Sequential Imputations and Bayesian Missing Data Problems. JASA 89(425), 278–288. → Effective sample size $1/\sum w_s^2$, the degeneracy diagnostic that triggers a refit.

Douc, R. and Cappé, O. (2005). Comparison of Resampling Schemes for Particle Filtering. ISPA. → Systematic resampling, chosen over multinomial for determinism.

Stephens, M. (2000). Dealing with Label Switching in Mixture Models. JRSS B 62(4), 795–809. → Why per-component summaries of a mixture posterior need post-hoc alignment.

Probabilistic programming

Goodman, N. D. and Stuhlmüller, A. (2014). The Design and Implementation of Probabilistic Programming Languages. dippl.org. → The model-as-program framing that fugue implements and that makes the grammar a prior rather than a generator function.

Ritchie, D., Horsfall, P. and Goodman, N. D. (2016). Deep Amortized Inference for Probabilistic Generative Models. arXiv:1610.05735. → Context for what trace-based inference over structured programs makes possible.

Grammar-based genetic programming

Whigham, P. A. (1995). Grammatically-based Genetic Programming. Workshop on Genetic Programming. → Using a grammar to constrain the search space so every individual is valid: Auracle's representation decision, with types in place of production rules.

Koza, J. R. (1992). Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press. → Tree-based GP, subtree crossover, and the bloat problem that a prior rather than a penalty addresses.

Takagi, H. (2001). Interactive Evolutionary Computation: Fusion of the Capabilities of EC Optimization and Human Evaluation. Proc. IEEE 89(9), 1275–1296. → The canonical statement of interactive evolution's user-fatigue bottleneck, which is the problem the two-loop architecture and the learned surrogate exist to solve.

Audio features and loudness

ITU-R BS.1770-4 (2015). Algorithms to measure audio programme loudness and true-peak audio level. → K-weighting, 400 ms gated blocks, the two gates. Implemented here

EBU R 128 (2020). Loudness normalisation and permitted maximum level of audio signals. → The practice around BS.1770 that makes −18 LUFS a sensible target.

Peeters, G. (2004). A large set of audio features for sound description. CUIDADO project report, IRCAM. → Spectral centroid, spread, flatness, rolloff and flux, in the definitions φ_audio uses.

Bregman, A. S. (1990). Auditory Scene Analysis. MIT Press. → Background for why octave-based frequency axes and segment-local measurements are the right coordinates for a perceptual feature vector.

Statistics of the feature space

Belsley, D. A., Kuh, E. and Welsch, R. E. (1980). Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. Wiley. → Variance inflation factors, the diagnostic that found two exact dependencies in φ_struct.

Huber, P. J. (1981). Robust Statistics. Wiley. → Winsorizing, and the reasoning behind using it as a fault detector rather than routinely.

The libraries

• quiver — github.com/alexnodeland/quiver · docs.rs
• fugue-evo — github.com/alexnodeland/fugue-evo · docs.rs
• fugue-ppl — docs.rs

Lineage

Auracle is the third iteration of one idea, and the two before it are worth knowing about because what each lacked is what this one is for:

v0 had the interaction but no engine. v1 had an engine, but a naive one with no memory of the user.