regularisation.h

Go to the documentation of this file.
00001 /***************************************************************************
00002  *   Copyright (C) 2005-2011 LpzRobots development team                    *
00003  *    Georg Martius  <georg dot martius at web dot de>                     *
00004  *    Frank Guettler <guettler at informatik dot uni-leipzig dot de        *
00005  *    Frank Hesse    <frank at nld dot ds dot mpg dot de>                  *
00006  *    Ralf Der       <ralfder at mis dot mpg dot de>                       *
00007  *                                                                         *
00008  *   This program is free software; you can redistribute it and/or modify  *
00009  *   it under the terms of the GNU General Public License as published by  *
00010  *   the Free Software Foundation; either version 2 of the License, or     *
00011  *   (at your option) any later version.                                   *
00012  *                                                                         *
00013  *   This program is distributed in the hope that it will be useful,       *
00014  *   but WITHOUT ANY WARRANTY; without even the implied warranty of        *
00015  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the         *
00016  *   GNU General Public License for more details.                          *
00017  *                                                                         *
00018  *   You should have received a copy of the GNU General Public License     *
00019  *   along with this program; if not, write to the                         *
00020  *   Free Software Foundation, Inc.,                                       *
00021  *   59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.             *
00022  *                                                                         *
00023  ***************************************************************************/
00024 #ifndef __REGULARISATION_H
00025 #define __REGULARISATION_H
00026 
00027 #include <cmath>
00028 #include <selforg/controller_misc.h>
00029 
00030 double inline sqr(double x) { 
00031   return x*x; 
00032 }
00033 
00034 /// neuron transfer function
00035 double inline g(double z)
00036 {
00037   return tanh(z);
00038 };
00039 
00040 /// first dervative
00041 double inline g_s(double z)
00042 {
00043   double k=tanh(z); 
00044   return 1.025 - k*k; 
00045   //  return 1/((1+0.5 * z*z)*(1+0.5 * z*z));    // softer
00046   //return 1/(1+log(1+z*z)); // even softer
00047 };
00048 
00049 
00050 /// first dervative with smoothing for large z
00051 double inline g_derivative(double z)
00052 {  
00053    return 1/((1+0.5 * z*z)*(1+0.5 * z*z));  
00054 };
00055 
00056 /// inverse of the first derivative
00057 double inline g_s_inv(double z)
00058 {
00059   double k=tanh(z);
00060    return 1/(1.025 - k*k);
00061   // return 1+z*z; // softer
00062   //return 1+log(1+z*z); // even softer
00063 };
00064 
00065 /** \f[ g'(z+xsi) = 1-(tanh(z+xsi))^2 \f] with additional clipping */
00066 double inline g_s(double z, double xsi) {
00067   double Z = clip(z, -3.0, 3.0) + clip(xsi, -1.0, 1.0);  
00068   double k=tanh(Z);  // approximation with Mittelwertsatz 
00069   return 1 - k*k;
00070 };
00071 
00072 
00073 /** soft version: \f[ g'(z+xsi) = 1/(1+(z+xsi)^2 \f] with additional clipping */
00074 double inline g_s_soft(double z, double xsi) {
00075   double Z = clip(z, -3.0, 3.0) + clip(xsi, -1.0, 1.0);  
00076   return 1/(1 + Z*Z);//TEST
00077 };
00078 
00079 
00080 /// an exact formula for g''/g'= -2g(Z), with clipped Z = z+xsi
00081 double inline g_ss_div_s(double z, double xsi) { 
00082   // for consistency reasons we use the same clipped z as for g'.
00083   double Z = clip(z, -3.0, 3.0) + clip(xsi, -1.0, 1.0);  
00084   // approximation with Mittelwertsatz (z is clipped)
00085   return -2*g(Z);
00086 };
00087 
00088 /// an soft formula for g''/g' = -2Z, with clipped Z = z+xsi
00089 double inline g_ss_div_s_soft(double z, double xsi) { 
00090   // for consistency reasons we use the same clipped z as for g'.
00091   double Z = clip(z, -3.0, 3.0) + clip(xsi, -1.0, 1.0);  
00092   return -2*Z;//TEST
00093 };
00094 
00095 /** with \f[ g'(z) = 1-(g(z+\xi))^2 \f] we get 
00096     \f[\frac{\partial}{\partial z} \frac{1}{g'(Z)} = \frac{g''}{g'^2} \f] 
00097     again with clipped Z
00098  */
00099 double inline derive_g_s_inv_exact_clip(double z, double xsi){
00100   double Z = clip(z, -3.0, 3.0) + clip(xsi, -1.0, 1.0);  
00101   double k=tanh(Z);  // approximation with Mittelwertsatz 
00102   return -2*k/(1-k*k);
00103 }
00104 
00105 /** \f[ g'(z) = 1-(z+\xi)^2 \f] which is the series expansion to the second order
00106 */ 
00107 double inline g_s_expand2(double z, double xsi){
00108   double Z = z + clip(xsi, -fabs(z), fabs(z));  
00109   return 1/(1+sqr(Z));
00110 }
00111 
00112 /** \f[ \frac{1}{g'(z)} \approx 1+(z+\xi)^2 \f] with geometric series approximation
00113 */ 
00114 double inline g_s_inv_expand2(double z, double xsi){
00115   double Z = z + clip(xsi, -fabs(z)/2.0, fabs(z)/2.0);  
00116   return 1+sqr(Z);
00117 }
00118 
00119 /** \f[ \frac{g''(z)}{g'(z)} \approx 2(z+\xi)(1+(z+\xi)^2) \f] with geometric series approximation
00120 */ 
00121 double inline g_ss_div_s_expand2(double z, double xsi){
00122   double Z = z + clip(xsi, -fabs(z)/2.0, fabs(z)/2.0);  
00123   //  double Z = z + clip(xsi, -fabs(z), fabs(z));  
00124   return -2*tanh(Z); 
00125 }
00126 
00127 
00128   /// squashing function (-0.1 to 0.1), to protect against to large weight updates
00129 double inline squash(double z)
00130   {
00131     return clip(z, -0.1, 0.1);
00132     //return 0.1 * tanh(10.0 * z);
00133   };
00134  
00135   /// squashing function with adjustable clipping size, to protect against too large weight updates
00136 double inline squash(void* d, double z) {
00137     double size = *((double*)d);
00138     return clip(z, -size, size);
00139   };
00140 
00141 
00142 
00143 #endif
00144 
Generated on Thu Jun 28 14:45:37 2012 for Robot Simulator of the Robotics Group for Self-Organization of Control by  doxygen 1.6.3