Austin Schuh | 9049e20 | 2022-02-20 17:34:16 -0800 | [diff] [blame] | 1 | Lasso |
| 2 | ===== |
| 3 | |
| 4 | |
| 5 | Lasso is a well known technique for sparse linear regression. |
| 6 | It is obtained by adding an :math:`\ell_1` regularization term in the objective, |
| 7 | |
| 8 | .. math:: |
| 9 | \begin{array}{ll} |
| 10 | \mbox{minimize} & \frac{1}{2} \| Ax - b \|_2^2 + \gamma \| x \|_1 |
| 11 | \end{array} |
| 12 | |
| 13 | |
| 14 | where :math:`x \in \mathbf{R}^{n}` is the vector of parameters, :math:`A \in \mathbf{R}^{m \times n}` is the data matrix, and :math:`\gamma > 0` is the weighting parameter. |
| 15 | The problem has the following equivalent form, |
| 16 | |
| 17 | .. math:: |
| 18 | \begin{array}{ll} |
| 19 | \mbox{minimize} & \frac{1}{2} y^T y + \gamma \boldsymbol{1}^T t \\ |
| 20 | \mbox{subject to} & y = Ax - b \\ |
| 21 | & -t \le x \le t |
| 22 | \end{array} |
| 23 | |
| 24 | |
| 25 | In order to get a good trade-off between sparsity of the solution and quality of the linear fit, we solve the problem for varying weighting parameter :math:`\gamma`. |
| 26 | Since :math:`\gamma` enters only in the linear part of the objective function, we can reuse the matrix factorization and enable warm starting to reduce the computation time. |
| 27 | |
| 28 | |
| 29 | |
| 30 | Python |
| 31 | ------ |
| 32 | |
| 33 | .. code:: python |
| 34 | |
| 35 | import osqp |
| 36 | import numpy as np |
| 37 | import scipy as sp |
| 38 | from scipy import sparse |
| 39 | |
| 40 | # Generate problem data |
| 41 | sp.random.seed(1) |
| 42 | n = 10 |
| 43 | m = 1000 |
| 44 | Ad = sparse.random(m, n, density=0.5) |
| 45 | x_true = np.multiply((np.random.rand(n) > 0.8).astype(float), |
| 46 | np.random.randn(n)) / np.sqrt(n) |
| 47 | b = Ad.dot(x_true) + 0.5*np.random.randn(m) |
| 48 | gammas = np.linspace(1, 10, 11) |
| 49 | |
| 50 | # Auxiliary data |
| 51 | In = sparse.eye(n) |
| 52 | Im = sparse.eye(m) |
| 53 | On = sparse.csc_matrix((n, n)) |
| 54 | Onm = sparse.csc_matrix((n, m)) |
| 55 | |
| 56 | # OSQP data |
| 57 | P = sparse.block_diag([On, sparse.eye(m), On], format='csc') |
| 58 | q = np.zeros(2*n + m) |
| 59 | A = sparse.vstack([sparse.hstack([Ad, -Im, Onm.T]), |
| 60 | sparse.hstack([In, Onm, -In]), |
| 61 | sparse.hstack([In, Onm, In])], format='csc') |
| 62 | l = np.hstack([b, -np.inf * np.ones(n), np.zeros(n)]) |
| 63 | u = np.hstack([b, np.zeros(n), np.inf * np.ones(n)]) |
| 64 | |
| 65 | # Create an OSQP object |
| 66 | prob = osqp.OSQP() |
| 67 | |
| 68 | # Setup workspace |
| 69 | prob.setup(P, q, A, l, u, warm_start=True) |
| 70 | |
| 71 | # Solve problem for different values of gamma parameter |
| 72 | for gamma in gammas: |
| 73 | # Update linear cost |
| 74 | q_new = np.hstack([np.zeros(n+m), gamma*np.ones(n)]) |
| 75 | prob.update(q=q_new) |
| 76 | |
| 77 | # Solve |
| 78 | res = prob.solve() |
| 79 | |
| 80 | |
| 81 | Matlab |
| 82 | ------ |
| 83 | |
| 84 | .. code:: matlab |
| 85 | |
| 86 | % Generate problem data |
| 87 | rng(1) |
| 88 | n = 10; |
| 89 | m = 1000; |
| 90 | Ad = sprandn(m, n, 0.5); |
| 91 | x_true = (randn(n, 1) > 0.8) .* randn(n, 1) / sqrt(n); |
| 92 | b = Ad * x_true + 0.5 * randn(m, 1); |
| 93 | gammas = linspace(1, 10, 11); |
| 94 | |
| 95 | % OSQP data |
| 96 | P = blkdiag(sparse(n, n), speye(m), sparse(n, n)); |
| 97 | q = zeros(2*n+m, 1); |
| 98 | A = [Ad, -speye(m), sparse(m,n); |
| 99 | speye(n), sparse(n, m), -speye(n); |
| 100 | speye(n), sparse(n, m), speye(n);]; |
| 101 | l = [b; -inf*ones(n, 1); zeros(n, 1)]; |
| 102 | u = [b; zeros(n, 1); inf*ones(n, 1)]; |
| 103 | |
| 104 | % Create an OSQP object |
| 105 | prob = osqp; |
| 106 | |
| 107 | % Setup workspace |
| 108 | prob.setup(P, q, A, l, u, 'warm_start', true); |
| 109 | |
| 110 | % Solve problem for different values of gamma parameter |
| 111 | for i = 1 : length(gammas) |
| 112 | % Update linear cost |
| 113 | gamma = gammas(i); |
| 114 | q_new = [zeros(n+m,1); gamma*ones(n,1)]; |
| 115 | prob.update('q', q_new); |
| 116 | |
| 117 | % Solve |
| 118 | res = prob.solve(); |
| 119 | end |
| 120 | |
| 121 | |
| 122 | |
| 123 | CVXPY |
| 124 | ----- |
| 125 | |
| 126 | .. code:: python |
| 127 | |
| 128 | from cvxpy import * |
| 129 | import numpy as np |
| 130 | import scipy as sp |
| 131 | from scipy import sparse |
| 132 | |
| 133 | # Generate problem data |
| 134 | sp.random.seed(1) |
| 135 | n = 10 |
| 136 | m = 1000 |
| 137 | A = sparse.random(m, n, density=0.5) |
| 138 | x_true = np.multiply((np.random.rand(n) > 0.8).astype(float), |
| 139 | np.random.randn(n)) / np.sqrt(n) |
| 140 | b = A.dot(x_true) + 0.5*np.random.randn(m) |
| 141 | gammas = np.linspace(1, 10, 11) |
| 142 | |
| 143 | # Define problem |
| 144 | x = Variable(n) |
| 145 | gamma = Parameter(nonneg=True) |
| 146 | objective = 0.5*sum_squares(A*x - b) + gamma*norm1(x) |
| 147 | prob = Problem(Minimize(objective)) |
| 148 | |
| 149 | # Solve problem for different values of gamma parameter |
| 150 | for gamma_val in gammas: |
| 151 | gamma.value = gamma_val |
| 152 | prob.solve(solver=OSQP, warm_start=True) |
| 153 | |
| 154 | |
| 155 | YALMIP |
| 156 | ------ |
| 157 | |
| 158 | .. code:: matlab |
| 159 | |
| 160 | % Generate problem data |
| 161 | rng(1) |
| 162 | n = 10; |
| 163 | m = 1000; |
| 164 | A = sprandn(m, n, 0.5); |
| 165 | x_true = (randn(n, 1) > 0.8) .* randn(n, 1) / sqrt(n); |
| 166 | b = A * x_true + 0.5 * randn(m, 1); |
| 167 | gammas = linspace(1, 10, 11); |
| 168 | |
| 169 | % Define problem |
| 170 | x = sdpvar(n, 1); |
| 171 | gamma = sdpvar; |
| 172 | objective = 0.5*norm(A*x - b)^2 + gamma*norm(x,1); |
| 173 | |
| 174 | % Solve with OSQP |
| 175 | options = sdpsettings('solver', 'osqp'); |
| 176 | x_opt = optimizer([], objective, options, gamma, x); |
| 177 | |
| 178 | % Solve problem for different values of gamma parameter |
| 179 | for i = 1 : length(gammas) |
| 180 | x_opt(gammas(i)); |
| 181 | end |