RcppLbfgsblaze

This package provides an implementation of the L-BFGS algorithm based on Blaze for R and Rcpp. The L-BFGS algorithm is a popular optimization algorithm for unconstrained optimization problems. Blaze is a high-performance C++ math library for dense and sparse arithmetic. The package provides a simple interface to the L-BFGS algorithm and allows users to optimize their objective functions with Blaze vectors and matrices in R and Rcpp.

Installation

You can install:

• the latest development version from github with

  r install.packages("remotes") remotes::install_github("ChingChuan-Chen/RcppLbfgsblaze")

If you encounter a bug, please file a reproducible example on github.

Logistic Model Fitting Benchmark

You can refer to the file logisticBenchmark.R to find the code. Below code and corresponding results show that RcppLbfgsblaze provides a fast and efficient algorithm for logistic model fitting. In the benchmark, RcppLbfgsBlaze is only slower than RcppNumerical, but faster than others.

``` r source(system.file("examples", "logisticBenchmark.R", package = "RcppLbfgsBlaze"))

logistic model fitting benchmark for n = 10000, p = 100 and non-zero p = 6: nrep = 20

Unit: milliseconds

expr min lq mean median uq max neval

glm.fit 254.8283 259.26340 277.47036 263.97765 283.17805 336.0213 20

optim 65.6519 69.05245 73.98623 70.32615 71.59410 148.2856 20

optim_arma 16.1281 16.97425 22.05490 17.81580 18.27810 94.6341 20

glmnet 38.5915 39.07820 44.26353 39.85460 41.21905 114.7233 20

lbfgs 85.7390 88.19755 96.58998 92.39210 94.74080 165.2018 20

lbfgs_arma 20.7813 21.87255 23.06615 22.71525 23.43450 29.2101 20

RcppNumerical 8.5511 9.06755 9.81808 9.54100 9.92005 16.1381 20

RcppLbfgsBlaze 10.0476 10.53910 11.26658 11.16235 11.59385 14.2053 20

```

When the sample size and number of predictors increase, RcppLbfgsBlaze will be faster than RcppNumerical and others. It shows that RcppLbfgsBlaze provides relatively fast algorithm comparing to otehrs.

``` r print(do_bench(5e4, 500))

logistic model fitting benchmark for n = 50000, p = 500 and non-zero p = 6: nrep = 20

Unit: milliseconds

expr min lq mean median uq max neval

optim_arma 547.1768 556.3317 578.6516 576.7488 594.5491 632.7897 20

lbfgs_arma 1501.6494 1537.9691 1573.0044 1561.6261 1606.5876 1675.4529 20

RcppNumerical 250.6525 255.3796 263.2894 262.5797 269.0240 286.2630 20

RcppLbfgsBlaze 150.5987 154.8733 158.7338 156.8559 161.5664 173.0606 20

```

Above results are run on my desktop (i9-13900K, DDR5-4000 128GB).

Authors

Ching-Chuan Chen

License

MIT License