gp.hpp 12.9 KB
Newer Older
1
2
#ifndef LIMBO_MODEL_GP_HPP
#define LIMBO_MODEL_GP_HPP
3

4
#include <iostream>
5
#include <cassert>
6
#include <limits>
7
#include <vector>
8

9
10
11
12
#include <Eigen/Core>
#include <Eigen/LU>
#include <Eigen/Cholesky>

13
#include <limbo/model/gp/no_lf_opt.hpp>
14

15
namespace limbo {
16
    namespace model {
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
17
18
19
20
21
        /// @ingroup model
        /// A classic Gaussian process.
        /// It is parametrized by:
        /// - a mean function
        /// - [optionnal] an optimizer for the hyper-parameters
22
        template <typename Params, typename KernelFunction, typename MeanFunction, class HyperParamsOptimizer = gp::NoLFOpt<Params>>
23
        class GP {
24
        public:
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
	  /// useful because the model might be created before knowing anything about the process
	  GP() : _dim_in(-1), _dim_out(-1) {}
	  
	  /// useful because the model might be created  before having samples
	  GP(int dim_in, int dim_out)
	    : _dim_in(dim_in), _dim_out(dim_out), _kernel_function(dim_in), _mean_function(dim_out) {}
	  
	  /// Compute the GP from samples, observation, noise. [optionnal: blacklisted samples]. This call needs to be explicit!
	  void compute(const std::vector<Eigen::VectorXd>& samples,
		       const std::vector<Eigen::VectorXd>& observations, double noise,
		       const std::vector<Eigen::VectorXd>& bl_samples = std::vector<Eigen::VectorXd>())
	  {
	    //should be checked each time! not only the first time
	    assert(samples.size() != 0);
	    assert(observations.size() != 0);
	    assert(samples.size() == observations.size());
	    
	    if (_dim_in != samples[0].size()) {
	      _dim_in = samples[0].size();
	      _kernel_function = KernelFunction(_dim_in); // the cost of building a functor should be relatively low
	    }
	    
	    if (_dim_out != observations[0].size()) {
	      _dim_out = observations[0].size();
	      _mean_function = MeanFunction(_dim_out); // the cost of building a functor should be relatively low
	    }
	    
	    
	    bool incremental =false;
	    if(samples.size()==_samples.size() || samples.size() == _samples.size()+1) //only one additional sample, can probably be generalized
	      {
		incremental = true;
		for(size_t i=0; incremental && i<_samples.size();i++)
		  incremental=_samples[i].isApprox(samples[i], 1e-1);
		
	      } 
	    
	    _noise = noise;
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78

	    if(samples.size()!=_samples.size() || !incremental) //no changes in samples since last computation.
	      {
		_samples = samples;
		
		_observations.resize(observations.size(), _dim_out);
		for (int i = 0; i < _observations.rows(); ++i)
		  _observations.row(i) = observations[i];
		
		_mean_observation.resize(_dim_out);
		for (int i = 0; i < _observations.cols(); i++)
		  _mean_observation(i) = _observations.col(i).sum() / _observations.rows();
		
		_compute_obs_mean();
		_compute_kernel(incremental);
	      }
79
80
	    
	    _bl_samples = bl_samples;
81
82
	    if (_bl_samples.size() != 0)
	      _compute_bl_kernel();
83
84
85
86
87
	    
	    HyperParamsOptimizer()(*this);
	  }
	  
	  /**
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
88
             \rst
89
             return :math:`\mu`, :math:`\sigma^2` (unormalized). If there is no sample, return the value according to the mean function. Using this method instead of separate calls to mu() and sigma() is more efficient because some computations are shared between mu() and sigma().
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
90
             \endrst
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
	  */
	  std::tuple<Eigen::VectorXd, double> query(const Eigen::VectorXd& v) const
	  {
	    if (_samples.size() == 0 && _bl_samples.size() == 0)
	      return std::make_tuple(_mean_function(v, *this),
				     sqrt(_kernel_function(v, v)));
	    
	    if (_samples.size() == 0)
	      return std::make_tuple(_mean_function(v, *this),
				     _sigma(v, _compute_k_bl(v, _compute_k(v))));
	    
	    Eigen::VectorXd k = _compute_k(v);
	    return std::make_tuple(_mu(v, k), _sigma(v, _compute_k_bl(v, k)));
	  }
	  
	  /**
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
107
             \rst
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
108
             return :math:`\mu` (unormalized). If there is no sample, return the value according to the mean function.
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
109
             \endrst
110
111
112
113
114
115
116
117
118
	  */
	  Eigen::VectorXd mu(const Eigen::VectorXd& v) const
	  {
	    if (_samples.size() == 0)
	      return _mean_function(v, *this);
	    return _mu(v, _compute_k(v));
	  }
	  
	  /**
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
119
             \rst
120
             return :math:`\sigma^2` (unormalized). If there is no sample, return the value according to the mean function.
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
121
             \endrst
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
	  */
	  double sigma(const Eigen::VectorXd& v) const
	  {
	    if (_samples.size() == 0 && _bl_samples.size() == 0)
	      return sqrt(_kernel_function(v, v));
	    return _sigma(v, _compute_k_bl(v, _compute_k(v)));
	  }
	  
	  /// return the number of dimensions of the input
	  int dim_in() const
	  {
	    assert(_dim_in != -1); // need to compute first !
	    return _dim_in;
	  }
	  
	  /// return the number of dimensions of the output
	  int dim_out() const
	  {
	    assert(_dim_out != -1); // need to compute first !
	    return _dim_out;
	  }
	  
	  const KernelFunction& kernel_function() const { return _kernel_function; }
	  
	  KernelFunction& kernel_function() { return _kernel_function; }
	  
	  const MeanFunction& mean_function() const { return _mean_function; }
	  
	  MeanFunction& mean_function() { return _mean_function; }
	  
Jean-Baptiste Mouret's avatar
Jean-Baptiste Mouret committed
152
            /// return the maximum observation (only call this if the output of the GP is of dimension 1)
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
	     Eigen::VectorXd max_observation() const
	     {
		 if (_observations.cols() > 1)
		     std::cout << "WARNING max_observation with multi dimensional "
				  "observations doesn't make sense" << std::endl;
		 return _observations.maxCoeff();
	     }

	     /// return the mean observation (only call this if the output of the GP is of dimension 1)
	     Eigen::VectorXd mean_observation() const
	     {
		 // TO-DO: Check if _dim_out is correct?!
		 return _samples.size() > 0 ? _mean_observation
					    : Eigen::VectorXd::Zero(_dim_out);
	     }

	     const Eigen::MatrixXd& mean_vector() const { return _mean_vector; }

	     const Eigen::MatrixXd& obs_mean() const { return _obs_mean; }

	     /// return the number of samples used to compute the GP
	     int nb_samples() const { return _samples.size(); }

	     /** return the number of blacklisted samples used to compute the GP
	     \rst
	     For the blacklist concept, see the Limbo-specific concept guide.
	     \endrst
	     */
	     int nb_bl_samples() const { return _bl_samples.size(); }

	     /// update the GP
	     void update(bool incremental =false)
	     {
		 this->_compute_obs_mean(); // ORDER MATTERS
		 this->_compute_kernel(incremental);
	     }

	     /// return the likelihood (do not compute it!)
	     double get_lik() const { return _lik; }

	     /// set the likelihood (you need to compute it from outside!)
	     void set_lik(const double& lik) { _lik = lik; }

	     /// LLT matrix (from Cholesky decomposition)
	     //const Eigen::LLT<Eigen::MatrixXd>& llt() const { return _llt; }
	   const Eigen::MatrixXd& matrixL() const { return _matrixL; }

	     const Eigen::MatrixXd& alpha() const { return _alpha; }

	     /// return the list of samples that have been tested so far
	     const std::vector<Eigen::VectorXd>& samples() const { return _samples; }

	 protected:
	     int _dim_in;
	     int _dim_out;

	     KernelFunction _kernel_function;
	     MeanFunction _mean_function;

	     std::vector<Eigen::VectorXd> _samples;
	     Eigen::MatrixXd _observations;
	     std::vector<Eigen::VectorXd> _bl_samples; // black listed samples
	     Eigen::MatrixXd _mean_vector;
	     Eigen::MatrixXd _obs_mean;

	     double _noise;
	     Eigen::MatrixXd _alpha;
	     Eigen::VectorXd _mean_observation;

	     Eigen::MatrixXd _kernel;
	     // Eigen::MatrixXd _inverted_kernel;
	     Eigen::MatrixXd _l_matrix;

	   Eigen::MatrixXd _matrixL;
	   //Eigen::TriangularView<Eigen::MatrixXd, Eigen::Lower> _matrixL_view;
	   //Eigen::LLT<Eigen::MatrixXd>::Traits::MatrixL  _matrixL_t;
	     Eigen::MatrixXd _inv_bl_kernel;

	     double _lik;

	     void _compute_obs_mean()
	     {
		 _mean_vector.resize(_samples.size(), _dim_out);
		 for (int i = 0; i < _mean_vector.rows(); i++)
		     _mean_vector.row(i) = _mean_function(_samples[i], *this);
		 _obs_mean = _observations - _mean_vector;
	     }

	   void _compute_kernel(bool incremental )
	   {
	       // Check if new kernel is the same as before
	       if(!incremental)//_kernel.cols()==0 || !kernel.block(0,0,_kernel.rows(),_kernel.cols()).isApprox(_kernel, 1e-10)) // incremental LLT is impossible
		 {
		   _kernel.resize(_samples.size(), _samples.size());
		   // O(n^2) [should be negligible]
		   for (size_t i = 0; i < _samples.size(); i++)
		     for (size_t j = 0; j < _samples.size(); ++j)
		       _kernel(i, j) = _kernel_function(_samples[i], _samples[j]) + ((i == j) ? _noise : 0); // noise only on the diagonal

		   // O(n^3)
		   //  _inverted_kernel = _kernel.inverse();
		   _matrixL = Eigen::LLT<Eigen::MatrixXd>(_kernel).matrixL();
		 }
	       else //incremental LLT
		 //This part of the code is inpired from the Bayesopt Library (cholesky_add_row function). However, the mathematical fundations can be easily retrieved by detailling the equations of the extended L matrix that produces the desired kernel.
		 {
		   size_t n=_samples.size();
		   _kernel.conservativeResize(n, n);

		   for (size_t i = 0; i < n; ++i){
		     _kernel(i, n-1) = _kernel_function(_samples[i], _samples[n-1]) + ((i == n-1) ? _noise : 0); // noise only on the diagonal
		     _kernel(n-1, i) = _kernel(i, n-1);
		   }

		   _matrixL.conservativeResizeLike(Eigen::MatrixXd::Zero(n,n));
		   double L_j;
		   for (size_t j = 0; j < n-1; ++j)
		     {
		       L_j = _kernel(n-1,j) - (_matrixL.block(j,0,1,j) * _matrixL.block(n-1,0,1,j).transpose())(0,0);
		       _matrixL(n-1,j) = (L_j) / _matrixL(j,j);
		     }

		   L_j = _kernel(n-1,n-1) - (_matrixL.block(n-1,0,1,n-1) * _matrixL.block(n-1,0,1,n-1).transpose())(0,0);
		   _matrixL(n-1,n-1) = sqrt(L_j);

		 }

	       // alpha = K^{-1} * this->_obs_mean;
	       _alpha = _matrixL.template triangularView<Eigen::Lower>().solve(_obs_mean);
282
	       _matrixL.template triangularView<Eigen::Lower>().adjoint().solveInPlace(_alpha); //can probably be improved by avoiding to generate the view twice
283
		 
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
	   }

	  void _compute_bl_kernel()
	  {
	    Eigen::MatrixXd A1 = Eigen::MatrixXd::Identity(this->_samples.size(), this->_samples.size());
	    _matrixL.template triangularView<Eigen::Lower>().solveInPlace(A1);
	    _matrixL.template triangularView<Eigen::Lower>().transpose().solveInPlace(A1);
	    
	    _inv_bl_kernel.resize(_samples.size() + _bl_samples.size(),
				  _samples.size() + _bl_samples.size());
	    
	    Eigen::MatrixXd B(_samples.size(), _bl_samples.size());
	    for (size_t i = 0; i < _samples.size(); i++)
	      for (size_t j = 0; j < _bl_samples.size(); ++j)
		B(i, j) = _kernel_function(_samples[i], _bl_samples[j]);
	    
	    Eigen::MatrixXd D(_bl_samples.size(), _bl_samples.size());
	    for (size_t i = 0; i < _bl_samples.size(); i++)
	      for (size_t j = 0; j < _bl_samples.size(); ++j)
		D(i, j) = _kernel_function(_bl_samples[i], _bl_samples[j]) + ((i == j) ? _noise : 0);
	    
	    Eigen::MatrixXd comA = (D - B.transpose() * A1 * B);
	    Eigen::LLT<Eigen::MatrixXd> llt_bl(comA);
	    Eigen::MatrixXd comA1 = Eigen::MatrixXd::Identity(_bl_samples.size(), _bl_samples.size());
	    llt_bl.matrixL().solveInPlace(comA1);
	    llt_bl.matrixL().transpose().solveInPlace(comA1);
	    
	    // fill the matrix block wise
	    _inv_bl_kernel.block(0, 0, _samples.size(), _samples.size()) = A1 + A1 * B * comA1 * B.transpose() * A1;
	    _inv_bl_kernel.block(0, _samples.size(), _samples.size(),
				 _bl_samples.size()) = -A1 * B * comA1;
	    _inv_bl_kernel.block(_samples.size(), 0, _bl_samples.size(),
				 _samples.size()) = _inv_bl_kernel.block(0, _samples.size(), _samples.size(),
									 _bl_samples.size()).transpose();
	    _inv_bl_kernel.block(_samples.size(), _samples.size(), _bl_samples.size(),
				 _bl_samples.size()) = comA1;
	  }
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365

	     Eigen::VectorXd _mu(const Eigen::VectorXd& v, const Eigen::VectorXd& k) const
	     {
		 return (k.transpose() * _alpha) + _mean_function(v, *this).transpose();
	     }

	     double _sigma(const Eigen::VectorXd& v, const Eigen::VectorXd& k) const
	     {
		 double res;
		 if (_bl_samples.size() == 0) {
		   Eigen::VectorXd z = _matrixL.triangularView<Eigen::Lower>().solve(k);
		     res = _kernel_function(v, v) - z.dot(z);
		 }
		 else {
		     res = _kernel_function(v, v) - k.transpose() * _inv_bl_kernel * k;
		 }

		 return (res <= std::numeric_limits<double>::epsilon()) ? 0 : res;
	     }

	     Eigen::VectorXd _compute_k(const Eigen::VectorXd& v) const
	     {
		 Eigen::VectorXd k(_samples.size());
		 for (int i = 0; i < k.size(); i++)
		     k[i] = _kernel_function(_samples[i], v);
		 return k;
	     }

	     Eigen::VectorXd _compute_k_bl(const Eigen::VectorXd& v,
		 const Eigen::VectorXd& k) const
	     {
		 if (_bl_samples.size() == 0) {
		     return k;
		 }

		 Eigen::VectorXd k_bl(_samples.size() + _bl_samples.size());

		 k_bl.head(_samples.size()) = k;
		 for (size_t i = 0; i < _bl_samples.size(); i++)
		     k_bl[i + this->_samples.size()] = this->_kernel_function(_bl_samples[i], v);
		 return k_bl;
	     }
	 };
     }
 }
366

367
#endif