平面网格世界上的 Qlearning附Matlab代码.rar

clear, clc, close all % Agent learns to navigate across a gridworld with obstacles including a % large gap traversed by a bridge using Qlearning. % initialize env Im = double(imread('gridworld-bridge.png')); Im = imresize(rgb2gray(Im), [32 24]); GR = imquantize(Im, multithresh(Im, 1)); [m, n] = size(GR); start = sub2ind(size(GR), 31 * ones(1, 5), 6 : 10); finish = sub2ind(size(GR), 2 * ones(1, 5), 15 : 19); GR([start; finish]) = 1.5; imagesc(GR), colormap(gray), axis equal % initialize agent (row, col[, action]) Q = zeros(m, n, 4); Qsave = Q; target_policy = zeros(m, n); Action_Set = [1 0; 0 -1; -1 0; 0 1]'; alpha = 0.1; gamma = 0.95; tolerance = 1 / 1000; count = 0; learner_converging = true; while learner_converging [row, col] = ind2sub(size(GR), datasample(start, 1)); A = randi(4); step = Action_Set * (1 : 4 == A)'; SA = sub2ind(size(Q), row, col, A); [~, S2] = bound(GR, [row; col], step + [row; col]); reward = 0; % do initial backup Q(SA) = Q(SA) + alpha * (reward + gamma * max(Q(S2(1), S2(2), :)) - Q(SA)); episode_in_progress = true; while episode_in_progress row = S2(1); col = S2(2); A = randi(4); step = Action_Set * (1 : 4 == A)'; SA = sub2ind(size(Q), row, col, A); % get S' and R [episode_in_progress, S2] = bound(GR, [row; col], [row; col] + step); reward = ~episode_in_progress; % backup Q(SA) = Q(SA) + alpha * (reward + gamma * max(Q(S2(1), S2(2), :)) ... - Q(SA)); % improve target policy [~, target_policy(row, col)] = max(Q(row, col, :)); end count = count + 1; if mod(count, 1000) == 0 ee = max(max(max(Q - Qsave))); if ee <= tolerance fprintf('count %d\n', count) learner_converging = false; end Qsave = Q; end end