ROCK* - the algorithm in this program combines Gaussian kernel regression and natural gradient descent and is shown to have better convergence than FDSA, SPSA, IW-PGPE, REINFORCE, CMAES, AMalGaM, and CEM:
Jemin Hwangbo, Christian Gehring, Hannes Sommer, Roland Siegwart and Jonas Buchli.2014. ROCK* - Efficient Black-box Optimization for Policy Learning. Humanoids 2014, Madrid, Spain http://www.adrl.ethz.ch/archive/Humanoid2014_ROCKSTAR.pdf
The natural gradient is described here: S. Amari, and S.C. Douglas. 1998. Why natural gradient? IEEE 1213-1216 and examined in: James Martens.2015. New perspectives on the natural gradient method. http://arxiv.org/pdf/1412.1193v4.pdf
and the covariant matrix adaptation (CMA-ES-like) is from here: Nikolas Hansen and Andreas Ostermeier. 1996. Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. IEEE 312-317
The objective function is like the one presented in:
https://chessprogramming.wikispaces.com/Texel%27s+Tuning+Method
Coefficient a = 0.007 (in centipawns) from: Vladimir Medvedev. Определяем веса шахматньiх фигур регрессионньiм анализом (Determination of chess pieces weights by regression analysis). http://habrahabr.ru/post/254753/