Small optimizations.

Author: wwkong

src/frameworks/IAIPAL.m

@@ -191,29 +191,13 @@
     params_acg.t_start = t_start;
 
     % Set the correct stepsizes.
-    if strcmp(params.acg_steptype, 'constant')
-      params_acg.L_est = M_s;
-    elseif strcmp(params.acg_steptype, 'variable') 
-      if (i_first_acg_run)
-        o_z0 = copy(oracle);
-        o_z0.eval(z0);
-        iter = iter + 1;
-        o_z1 = copy(oracle);
-        o_z1.eval(z0 - o_z0.grad_f_s() / M);
-        iter = iter + 1;
-        z1 = o_z1.prox_f_n(1 / M);
-        o_z1.eval(z1);
-        iter = iter + 1;
-        if (norm_fn(z0 - z1) <= 1e-6)
-          params_acg.L_est = lambda * M + 1;
-        else
-          M_est = norm_fn(o_z0.grad_f_s() - o_z1.grad_f_s()) / norm_fn(z0 - z1);
-          params_acg.L_est = lambda * M_est + 1;
-        end
-      else
-        % Use the previous estimate.
-        params_acg.L_est = model_acg.L_est;
-      end
+    if ( strcmp(params.acg_steptype, 'constant'))
+        params_acg.L_est = M_s;
+    elseif (i_first_acg_run && strcmp(params.acg_steptype, 'variable'))
+      params_acg.L_est = lambda * M + 1;
+    elseif (~i_first_acg_run && strcmp(params.acg_steptype, 'variable'))
+      % Use the previous estimate.
+      params_acg.L_est = model_acg.L_est;
     else
       error('Unknown ACG steptype!');
     end
@@ -270,7 +254,7 @@
       history.norm_p_hat = [history.norm_p_hat; norm_fn(p_hat)];
       history.norm_v = [history.norm_v; norm_fn(w_hat)];
     end
-    
+        
     % Check for termination.
     if (isempty(params.termination_fn) || params.check_all_terminations)
       if (norm_fn(w_hat) <= opt_tol && norm_fn(q_hat) <= feas_tol)

src/frameworks/penalty.m

@@ -25,9 +25,6 @@
 %   A pair of structs containing model and history related outputs of the solved problem associated with the oracle and input
 %   parameters.
 
-  % Global constants.
-  MIN_PENALTY_CONST = 1;
-
   % Timer start.
   t_start = tic;
 
@@ -89,7 +86,7 @@
   stage = 1;
   solver_params = params;
   i_reset_prox_center = params.i_reset_prox_center;
-  c = max([MIN_PENALTY_CONST, params.M / params.K_constr ^ 2]);
+  c = params.c0;
 
   %% MAIN ALGORITHM
   while true
@@ -146,9 +143,16 @@
     end
 
     % Check for termination.
-    feas = feas_fn(solver_model.x);
-    if (feas <= feas_tol)
-      break;
+    feas = feas_fn(solver_model.x);  
+    if (~isempty(params.termination_fn) || params.check_all_terminations)
+      if (feas <= feas_tol)
+        break;
+      end
+    end
+    if (~isempty(params.termination_fn) || params.check_all_terminations)
+      if wrap_termination(params, solver_model.x, c)
+        break;
+      end
     end
 
     % Apply warm-start onto the initial point fed into the next iteration.
@@ -164,12 +168,11 @@
 
   % Prepare to output
   model = solver_model;
-  if (isempty(params.termination_fn) || params.check_all_terminations)
+  if (isempty(params.termination_fn))
     [~, feas_vec] = feas_fn(solver_model.x);
     model.w = feas_vec;
-  end
-  if (~isempty(params.termination_fn) || params.check_all_terminations)
-    [~, model.v, model.w] = wrap_termination(params, solver_model.x, c);  
+  else
+    [~, model.v, model.w] = wrap_termination(params, solver_model.x, c);
   end
   history.iter = iter;
   history.stage = stage;
@@ -187,6 +190,10 @@
 
 % Fills in parameters that were not set as input.
 function params = set_default_params(params)
+
+  % Global constants.
+  MIN_PENALTY_CONST = 1;
+
   % Overwrite if necessary.
   if (~isfield(params, 'i_logging')) 
     params.i_logging = false;
@@ -206,4 +213,7 @@
   if (~isfield(params, 'check_all_terminations'))
     params.check_all_terminations = false;
   end
+  if (~isfield(params, 'c0'))
+    params.c0 = max([MIN_PENALTY_CONST, params.M / params.K_constr ^ 2]);
+  end
 end

tests/papers/nl_iapial/extra/nl_iapial_qsdp_debug.m

@@ -4,9 +4,9 @@
 
 n = 100;
 r = 1;
-m = 1e4;
-M = 1e6;
-global_tol = 1e-2;
+m = 1e0;
+M = 1e4;
+global_tol = 1e-4;
 
 % Run an instance via the command line.
 print_tbls(n, r, m, M, global_tol);
@@ -71,20 +71,24 @@ function run_experiment(N, r, M, m, dimM, dimN, density, seed, global_tol)
   base_hparam.check_all_terminations = true;
 
   % Create the IAPIAL hparams.
-  qp_hparam = base_hparam;
-  qp_hparam.acg_steptype = 'constant';
-  qp_hparam.aipp_type = 'aipp';
+  qpa_hparam = base_hparam;
+  qpa_hparam.acg_steptype = 'variable';
+  qpa_hparam.aipp_type = 'aipp';
   ipl_hparam = base_hparam;
   ipl_hparam.acg_steptype = 'constant';
+  ipl_hparam.sigma_min = sqrt(0.3);
+  ipla_hparam = base_hparam;
+  ipla_hparam.acg_steptype = 'variable';
+  ipla_hparam.sigma_min = sqrt(0.3);
 
   % Run a benchmark test and print the summary.
-  hparam_arr = {qp_hparam, ipl_hparam};
-  name_arr = {'QP','IPL'};
+  hparam_arr = {qpa_hparam, ipla_hparam};
+  name_arr = {'QP_A', 'IPL_A'};
   framework_arr = {@penalty, @IAIPAL};
   solver_arr = {@AIPP, @ECG};
 
   % Run the test.
-  [~, summary_mdls] = run_CCM_benchmark(ncvx_qsdp, framework_arr, solver_arr, hparam_arr, name_arr);
+  [summary_tbls, summary_mdls] = run_CCM_benchmark(ncvx_qsdp, framework_arr, solver_arr, hparam_arr, name_arr);
   for s=name_arr
     disp(['Algorithm = ', s{1}]);
     disp(table(dimN, m, M, r));
@@ -93,5 +97,6 @@ function run_experiment(N, r, M, m, dimM, dimN, density, seed, global_tol)
     tbl.Properties.VariableNames = {'inner_iter', 'L_est', 'norm_v'};
     disp(tbl);
   end
+  disp(summary_tbls.all);
 
 end

tests/papers/nl_iapial/nl_iapial_qcqp.m

@@ -101,7 +101,7 @@ function print_tbls(dimN)
   % Set the tolerances
   ncvx_qc_qp.opt_tol = global_tol;
   ncvx_qc_qp.feas_tol = global_tol;
-  ncvx_qc_qp.time_limit = 6000;
+  ncvx_qc_qp.time_limit = 3000;
   ncvx_qc_qp.iter_limit = 1000000;
 
   % Add linear constraints
@@ -122,11 +122,10 @@ function print_tbls(dimN)
   % Create the IAPIAL hparams.
   ipl_hparam = base_hparam;
   ipl_hparam.acg_steptype = 'constant';
-  ipl_hparam.sigma_min = 1/sqrt(2);
+  ipl_hparam.sigma_min = sqrt(0.3);
   ipla_hparam = base_hparam;
   ipla_hparam.acg_steptype = 'variable';
-  ipla_hparam.init_mult_L = 0.5;
-  ipla_hparam.sigma_min = 1/sqrt(2);
+  ipla_hparam.sigma_min = sqrt(0.3);
 
   % Run a benchmark test and print the summary.
   hparam_arr = {ipl_hparam, ipla_hparam};

tests/papers/nl_iapial/nl_iapial_qcqsdp.m

@@ -64,7 +64,7 @@ function print_tbls(dimN)
   disp(o_tbl);
 
 end
-function o_tbl =   run_experiment(N, r, M, m, dimM, dimN, density, seed, global_tol)
+function o_tbl = run_experiment(N, r, M, m, dimM, dimN, density, seed, global_tol)
 
   [oracle, hparams] = test_fn_quad_cone_constr_02r(N, r, M, m, seed, dimM, dimN, density);
 
@@ -118,8 +118,10 @@ function print_tbls(dimN)
   % Create the IAPIAL hparams.
   ipl_hparam = base_hparam;
   ipl_hparam.acg_steptype = 'constant';
+  ipl_hparam.sigma_min = sqrt(0.3);
   ipla_hparam = base_hparam;
   ipla_hparam.acg_steptype = 'variable';
+  ipla_hparam.sigma_min = sqrt(0.3);
 
   % Create the complicated iALM hparams.
   ialm_hparam = base_hparam;
@@ -129,7 +131,6 @@ function print_tbls(dimN)
   ialm_hparam.L0 = max([hparams.m, hparams.M]);
   ialm_hparam.rho_vec = hparams.m_constr_vec;
   ialm_hparam.L_vec = hparams.L_constr_vec;
-  % Note that we are using the fact that |X|_F <= 1 over the eigenbox.
   ialm_hparam.B_vec = hparams.K_constr_vec;
 
   % Run a benchmark test and print the summary.

tests/papers/nl_iapial/nl_iapial_qp.m

@@ -152,7 +152,7 @@ function print_tbls(dimN)
   % Set the tolerances
   ncvx_qsdp.opt_tol = global_tol;
   ncvx_qsdp.feas_tol = global_tol;
-  ncvx_qsdp.time_limit = 6000;
+  ncvx_qsdp.time_limit = 3000;
 
   % Add linear constraints
   ncvx_qsdp.constr_fn = hparams.constr_fn;
@@ -171,6 +171,9 @@ function print_tbls(dimN)
   % Create the IAPIAL hparams.
   ipla_hparam = base_hparam;
   ipla_hparam.acg_steptype = 'variable';
+  ipla_hparam.sigma_min = sqrt(0.3);
+  
+  % Create the QP-AIPP hparams.
   qpa_hparam = base_hparam;
   qpa_hparam.acg_steptype = 'variable';
   qpa_hparam.aipp_type = 'aipp';

tests/papers/nl_iapial/nl_iapial_qsdp.m

@@ -15,7 +15,7 @@ function print_tbls(dimN)
   dimM = 25;
   N = 1000;
   density = 0.05;
-  global_tol = 1e-3;
+  global_tol = 5 * 1e-4;
   m_vec = [1e2, 1e3, 1e4];
   M_vec = [1e4, 1e5, 1e6];
   r_vec = [5, 10, 20];
@@ -83,7 +83,8 @@ function print_tbls(dimN)
   rho = global_tol * (1 + hparams.norm_fn(o_at_x0.grad_f_s()));
   eta = global_tol * (1 + hparams.norm_fn(g0 - hparams.set_projector(g0)));
   alt_grad_constr_fn = @(x, p) tsr_mult(hparams.grad_constr_fn(x), p, 'dual');
-  term_wrap = @(x,p) termination_check(x, p, o_at_x0, hparams.constr_fn, alt_grad_constr_fn, @proj_dh, @proj_NKt, hparams.norm_fn, rho, eta);  
+  term_wrap = @(x,p) ...
+    termination_check(x, p, o_at_x0, hparams.constr_fn, alt_grad_constr_fn, @proj_dh, @proj_NKt, hparams.norm_fn, rho, eta);  
 
   % Create the Model object and specify the solver.
   ncvx_qsdp = ConstrCompModel(oracle);
@@ -94,10 +95,10 @@ function print_tbls(dimN)
   ncvx_qsdp.m = hparams.m;
   ncvx_qsdp.K_constr = hparams.K_constr;
 
-  % Set the tolerances
+  % Set the tolerances.
   ncvx_qsdp.opt_tol = global_tol;
   ncvx_qsdp.feas_tol = global_tol;
-  ncvx_qsdp.time_limit = 12000;
+  ncvx_qsdp.time_limit = 6000;
 
   % Add linear constraints
   ncvx_qsdp.constr_fn = hparams.constr_fn;
@@ -116,8 +117,12 @@ function print_tbls(dimN)
   % Create the IAPIAL hparams.
   ipl_hparam = base_hparam;
   ipl_hparam.acg_steptype = 'constant';
+  ipl_hparam.sigma_min = sqrt(0.3);
   ipla_hparam = base_hparam;
   ipla_hparam.acg_steptype = 'variable';
+  ipla_hparam.sigma_min = sqrt(0.3);
+  
+  % Create the QP-AIPP hparams.
   qp_hparam = base_hparam;
   qp_hparam.acg_steptype = 'constant';
   qp_hparam.aipp_type = 'aipp';
@@ -133,7 +138,6 @@ function print_tbls(dimN)
   ialm_hparam.L0 = max([hparams.m, hparams.M]);
   ialm_hparam.rho_vec = hparams.m_constr_vec;
   ialm_hparam.L_vec = hparams.L_constr_vec;
-  % Note that we are using the fact that |X|_F <= 1 over the eigenbox.
   ialm_hparam.B_vec = hparams.K_constr_vec;
 
   % Run a benchmark test and print the summary.