 |
    |
Natural Gradient Approach for Linearly Constrained Continuous Optimization
Youhei Akimoto1 and Shinichi Shirakawa2
1Faculty of Engineering, Shinshu University, Nagano, Nagano, Japan
y_akimoto@shinshu-u.ac.jp
2College of Science and Engineering, Aoyama Gakuin University, Sagamihara, Kanagawa, Japan
shirakawa@it.aoyama.ac.jp
Abstract. When a feasible set of an optimization problem is a proper subset of a multidimensional real space and the optimum of the problem is located on or near the boundary of the feasible set, most evolutionary algorithms require a constraint handling machinery to generate better candidate solutions in the feasible set. However, some standard constraint handling such as a resampling strategy affects the distribution of the candidate solutions; the distribution is truncated into the feasible set. Then, the statistical meaning of the update of the distribution parameters will change. To construct the parameter update rule for the covariance matrix adaptation evolution strategy from the same principle as unconstrained cases, namely the natural gradient principle, we derive the natural gradient of the log-likelihood of the Gaussian distribution truncated into a linearly constrained feasible set. We analyze the novel parameter update on a minimization of a spherical function with a linear constraint.
LNCS 8672, p. 252 ff.
lncs@springer.com
© Springer International Publishing Switzerland 2014