基于强化学习的自适应机器人控制算法实现.zip_matlab simulink_强化学习控制_机器人

function run_algorithm(task, cooplev, method, cfg) % RUN_ALGORITHM Runs a given MARL algorithm on a given type of task. % RUN_ALGORITHM(TASK, COOPLEV, METHOD, CFG) % % Runs a MARL algorithm on a given type of task. Uses 5x5 robots % gridworlds as tasks. Specifically designed for use with alg_gui, but can % also be used independently. Saves the algorithm data in a field of the % 'algdata' structure in the base workspace. Creates the variable if not % present. The field name is created as a combination of task type and % algorihtm type. Any previous data for that task/algorithm pair is % cleared and the resources associated with it released. % % Parameters: % TASK - specifies the type of task. One of 'nav', 'rescue', % for navigation task, or search & rescue task, respectively. % COOPLEV - specifies the cooperation level in the task. One of 'coop', 'comp', % 'mixed', for fully cooperative, fully competitive, and mixed tasks, % respectively. % METHOD - specifies the algorithm to run. One of 'q' (single-agent % Q-learning), 'fsq' (full-state Q-learning), 'asf', (adaptive state % focus Q-learning), 'teamq' (team Q-learning), 'wolf', % (Win-or-Learn-Fast policy hill climbing), 'opt' (predefined optimal % sequences on the given task). % CFG - configuration of learning process. Structure with the % following fields (all optional) % 'stopearly' - if learning should be stopped upon % convergence. Default 1 (yes). % 'showafter' - make world visible only after this number of % trials. If Inf, world will be invisible throughout the learnign % process. Default Inf. % 'slowdownafter' - similarly, but for slowing down the learning % process for visualization. Takes effect only after 'showafter' % trials. Default Inf. % 'pausebeforerun' - if world should be held on screen until % keypress before starting the learning. Default 1 (hold on % screen). % 'useseeds' - whether the rand generator should be % deterministically initialized for reproducible behavior of % algorithms. Default 1 (use reproducible behavior). % % See also replay_algorithm % Author: Lucian Busoniu % Version: 2.0 % History: 6 Sept 2006, 2.0 -- added support for robot rescue gridworld, removed competitive / % minimax code DEF_CFG.stopearly = 1; DEF_CFG.pausebeforerun = 1; DEF_CFG.useseeds = 1; DEF_VIEWCFG.repres = 'abstract'; % no longer present in GUI, always abstract repres DEF_VIEWCFG.showafter = Inf; DEF_VIEWCFG.slowdownafter = Inf; DEF_CFG.viewcfg = DEF_VIEWCFG; if nargin < 3, cfg = DEF_CFG; else cfg = checkparams(cfg, DEF_CFG); end; % task % cooplev % method % cfg % viewcfg = cfg.viewcfg % return % top-level def learning params Nag = 2; % two agents - most of the times lp.trials = 100; lp.maxiter = 1000; if cfg.stopearly, lp.convgoal = 0.001; else lp.convgoal = -1; end; lp.avgtrials = 10; lp.showconv = 1; % show online conv plot for all algorithms lp.closeconv = 0; % top-level default agent learning params alp.alpha = .1; alp.gamma = .98; alp.lambda = 0.5; alp.epsilon = .7; % init structs seeds = struct(); OVWR = struct(); % --------------------------- % By task type and cooperation level % setup task structure + reward function % setup common learning params % setup learning params overwrites switch task, % ------------------------------------------ % Navigation task case 'nav', % common learning parameters/overwrites for navigation task alp.window = 256; % for ASF lp.trials = 100; lp.maxiter = 1000; alp.alpha = .2; alp.gamma = .98; alp.lambda = 0.5; alp.epsilon = .7; % A set of optimal action sequences OPTSEQ.coop = {'ursruusu', 'ulluurul'}; OPTSEQ.mixed = {'ursruuurr', 'ulluuull'}; switch cooplev, case 'coop', grid = [1 0 4 0 1; % Y 0 0 0 0 0; % | 1 1 0 1 1; % | 0 0 0 0 0; % | 2 0 0 0 3]; % | % 0 -- -------> X 0 goal.type = 'coop'; goal.rew = [-.1 NaN 10]; % random seeds seeds.q = 1e12; seeds.wolf = 1e6; seeds.teamq = 1e3; seeds.fsq = 1e10; seeds.asf = 1e9; % max trials settings, -1 = default max_trials.q = 40; max_trials.wolf = -1; max_trials.teamq = -1; max_trials.fsq = -1; max_trials.asf = -1; case 'mixed', % agent 1: 2, agent 2: 3, goal 1: 4, goal 2: 5 grid = [5 0 0 0 4; % Y 0 0 0 0 0; % | 1 1 0 1 1; % | 0 0 0 0 0; % | 2 0 0 0 3]; % | % 0 -- -------> X 0 goal.type = 'mixed'; goal.rew = [-.1 -.1 10]; % random seeds seeds.q = 1e12; seeds.wolf = 1e6; % 15 steps, but can't get it to work better... seeds.teamq = 1e3; seeds.fsq = 1e10; seeds.asf = 1e9; % max trials settings, -1 = default max_trials.q = 40; max_trials.wolf = -1; max_trials.fsq = -1; max_trials.asf = -1; otherwise, error(['Unknown cooperation level [' cooplev ']']); end; % cooplev switch for navigation task grid = fliplr(grid'); % process grid matrix into world data structures [i j] = find(grid == 1); obstacles = [i j]'; worldsize = size(grid)'; [x1 y1] = find(grid == 2); [x2 y2] = find(grid == 3); [xg1 yg1] = find(grid == 4); [xg2 yg2] = find(grid == 5); positions = [x1 y1; x2 y2]'; goal.pos = [xg1 yg1; xg2 yg2]'; % ----------------------------------- % param overwrites for methods % custom rewards for teamq ovwr.name = 'goal.rew'; ovwr.value = '[0 NaN 10]'; OVWR.teamq = {ovwr}; % END navigation task CASE % ------------------------------------------ % Rescue robotics simulated task case 'rescue', % common lps/overwrites for the rescue task lp.avgtrials = 20; % common lps for algorithms on this task lp.trials = 100; lp.maxiter = 1000; alp.alpha = .2; alp.gamma = .98; alp.lambda = 0.5; alp.epsilon = .7; switch cooplev, case 'coop', % % Agents can pass narrowway, have to coordinate upon return to homebase % grid = [0 0 0 0 0 0; % 0 0 0 101.1 101 101.2; % | % 0 0 0 -1 -1 -1; % | % 0 0 0 0 0 2; % 1 10 10 10 -1 0]; % | % % 0 -- -------> X 0 % % random seeds % seeds.q = 1e10; % OK, stop near target % seeds.wolf = 0;