Algorithmic Differentiation: C++ & Extremum Estimation
summary
Describes using C++ algorithmic differentiation to compute exact gradients via the chain rule and plug them into a numerical optimizer to recover model parameters a=0.10 and b=0.85 from simulated data. Builds a 10,000‑run simulation and optimizer that yields 99.49% convergence versus about 3% for finite differences; finite differences show noisy, step‑size‑dependent errors while AD delivers fast, accurate extremum estimation and integrates with Eigen and open‑source AD libraries.
AI-written summary · about
chapters (9)
- 0:00Start
- 2:11MOTIVATION: PARAMETRIC MODELS
- 2:39NUMERICAL OPTIMIZATION: ALGORITHMS
- 4:40NUMERICAL OPTIMIZATION: ALGORITHMIC DIFFERENTIATION
- 5:31CALCULATING DERIVATIVES
- 7:49FLOATING POINT ARITHMETIC
- 11:51FINITE DIFFERENCE APPROXIMATION ERROR
- 12:22SYMBOLIC DIFFERENTIATION
- 12:35NUMERICAL OBJECTIVE FUNCTION & ALGORITHMIC DIFFERENTIATION
related talks
59mDifferentiable Programming in C++
46mGradient Descent: The Ultimate Optimizer
01h 18mOptimization I
29mPartial Evaluation of Automatic Differentiation for Differential-Algebraic Equati...
26mADEV: Sound Automatic Differentiation of Expected Values of Probabilistic Progra...
11mAbstract Interpretation for Automatic Differentiation
19mGetting started with automatic differentiation
01h 07mOptimization Crash Course (continued)