QPFit/analyze_fourstate.m
3bfb25b5e70amaster

analyze_fourstate.m
View Options

			% Setup that worked:

			param = dict()
			param['batch_size'] = 50
			param['nVar'] = 10
			param['maxIter'] = 20

			param['num_epochs'] = 4
			param['log_interval'] = 500
			param['learning_rate'] = 1e-3




			%% Learned model
			load data/fourstate_fit.mat

			fit.F = double(F);
			fit.f = double(f(:));
			fit.L = double(L);

			% min .5z'Qz + (Fx + f')'*z
			% s.t. z >= 0

			nVar = size(fit.L,2);
			nParam = size(fit.F,2);

			fit.Q = fit.L'fit.L + 1e-3eye(nVar);

			z = sdpvar(nVar,1);
			x = sdpvar(nParam,1);
			obj = 0.5z'fit.Qz + (fit.Fx+fit.f)'*z;

			g = sdpsettings('solver','gurobi+');
			g.gurobi.IterationLimit = double(maxIter);

			fit.opt= optimizer(z >= 0, obj, g, x, z);
			fit.prob = Opt([z >= 0; -ROI <= x <= ROI],obj,x,z);
			% fit.prob.bth = ROI * ones(4,1);
			% fit.sol = fit.prob.solve();

			fit.ctrl = @(x) min(5, max(-5, O*fit.opt(x)+o(:)));
			% ul = fit.ctrl(X);

			%%
			sys.A = [0.7 -0.1 0.0 0.0; 0.2 -0.5 0.1 0.0; 0.0 0.1 0.1 0.0; 0.5 0.0 0.5 0.5];
			sys.B = [0.0 0.1; 0.1 1.0; 0.1 0.0; 0.0 0.0];
			sys.C = [1 1 1 1];
			sys.D = [0 0];

			n = 4;
			m = 2;

			clf
			maxN = 20;

			for i = 1:10
			x0 = ROI * (2 * rand(n,1) - 1);
			[Y, X, U] = simsys(sys, x0, fit.ctrl, maxN);
			plot(Y,'b')
			hold on
			[Y, X, U] = simsys(sys, x0, gnd.opt, maxN);

			plot(Y,'r')
			end