import numpy as np import numpy.random as rnd import torch as pt import math from gym import spaces, logger from gym.utils import seeding from system import System import time class SwingUpEnv(): """ Description: A pole is attached by an un-actuated joint to a cart, which moves along a frictionless track. The pendulum starts upright, and the goal is to prevent it from falling over by increasing and reducing the cart's velocity. Source: This environment corresponds to the version of the cart-pole problem described by Barto, Sutton, and Anderson Observation: Type: Box(4) Num Observation Min Max 0 Cart Position -4.8 4.8 1 Cart Velocity -Inf Inf 2 Pole Angle -Inf Inf 3 Pole Velocity At Tip -Inf Inf Actions: Type: Box(1) Num Action Min Max 0 Push cart -1 1 Note: The amount the velocity that is reduced or increased is not fixed; it depends on the angle the pole is pointing. This is because the center of gravity of the pole increases the amount of energy needed to move the cart underneath it Reward: Reward is 1 for every step taken, including the termination step Starting State: All observations are assigned a uniform random value in [-0.05..0.05] Episode Termination: Pole Angle is more than 12 degrees Cart Position is more than 2.4 (center of the cart reaches the edge of the display) Episode length is greater than 200 Solved Requirements Considered solved when the average reward is greater than or equal to 195.0 over 100 consecutive trials. """ metadata = { 'render.modes': ['human', 'rgb_array'], 'video.frames_per_second' : 50 } def __init__(self): self.sys = System(angular_units='Radians') self.force_mag = 10.0 self.tau = 0.02 # seconds between state updates # Angle at which to fail the episode self.x_threshold = 10. self.x_dot_threshold = 10. self.theta_dot_threshold = 3*np.pi # Angle limit set to 2 * theta_threshold_radians so failing observation is still within bounds high = np.array([self.x_threshold*2, self.x_dot_threshold, np.finfo(np.float32).max, np.finfo(np.float32).max]) self.action_space = spaces.Box(-np.ones(1), np.ones(1), dtype = np.float32) self.seed() self.state = None self.steps_beyond_done = None def seed(self, seed=None): self.np_random, seed = seeding.np_random(seed) return [seed] def step(self, action): assert self.action_space.contains(action), "%r (%s) invalid"%(action, type(action)) state = self.state x, x_dot, theta, theta_dot = state force = self.force_mag * action[0] self.sys.adjust(force) costheta = math.cos(theta) sintheta = math.sin(theta) if costheta > 0: self.up_time += 1 self.max_up_time = np.max([self.up_time, self.max_up_time]) else: self.up_time = 0 new_theta, new_x = self.sys.measure() theta_dot = (new_theta - theta) / self.tau x_dot = (new_x - x) / self.tau self.state = (new_x, x_dot, new_theta, theta_dot) self.sys.add_results(new_theta, new_x, force) done = x < -self.x_threshold \ or x > self.x_threshold \ or theta_dot < -self.theta_dot_threshold \ or theta_dot > self.theta_dot_threshold done = bool(done) if not done: reward = np.ceil(costheta) elif self.steps_beyond_done is None: # Pole just fell! self.steps_beyond_done = 0 reward = -( 100 * (np.abs(x_dot) + np.abs(theta_dot)) ) else: if self.steps_beyond_done == 0: logger.warn("You are calling 'step()' even though this environment has already returned done = True. You should always call 'reset()' once you receive 'done = True' -- any further steps are undefined behavior.") self.steps_beyond_done += 1 reward = 0.0 return np.array(self.state), reward, done, {'max_up_time' : self.max_up_time} def reset(self): self.sys.return_home() time.sleep(1) self.state = (0, 0, np.pi, 0) self.up_time = 0 self.max_up_time = 0 self.up = False self.steps_beyond_done = None return np.array(self.state) class nnQ(pt.nn.Module): """ Here is a basic neural network with for representing a policy """ def __init__(self, stateDim, numActions, numHiddenUnits, numLayers): super().__init__() InputLayer = [pt.nn.Linear(stateDim + numActions, numHiddenUnits), pt.nn.ReLU()] HiddenLayers = [] for _ in range(numLayers - 1): HiddenLayers.append(pt.nn.Linear(numHiddenUnits, numHiddenUnits)) HiddenLayers.append(pt.nn.ReLU()) OutputLayer = [pt.nn.Linear(numHiddenUnits, 1)] AllLayers = InputLayer + HiddenLayers + OutputLayer self.net = pt.nn.Sequential(*AllLayers) self.numActions = numActions def forward(self,x,a): x = pt.tensor(x, dtype = pt.float32) b = pt.nn.functional.one_hot(pt.tensor(a).long(), self.numActions) c = b.float().detach() y = pt.cat([x, c]) return self.net(y) class sarsaAgent: def __init__(self, stateDim, numActions, numHiddenUnits, numLayers, epsilon = .1, gamma = .9, alpha = .1): self.Q = nnQ(stateDim, numActions, numHiddenUnits, numLayers) self.gamma = gamma self.epsilon = epsilon self.alpha = alpha self.numActions = numActions self.s_last = None def action(self, x): # This is an epsilon greedy selection a = 0 if rnd.rand() < self.epsilon: a = rnd.randint(numActions) else: qBest = -np.inf for aTest in range(self.numActions): qTest = self.Q(x, aTest).detach().numpy()[0] if qTest > qBest: qBest = qTest a = aTest return a def update(self, s, a, r, s_next,done): # Compute the TD error, if there is enough data update = True if done: Q_cur = self.Q(s, a).detach().numpy()[0] delta = r - Q_cur self.s_last = None Q_diff = self.Q(s, a) elif self.s_last is not None: Q_next = self.Q(s, a).detach().numpy()[0] Q_cur = self.Q(self.s_last, self.a_last).detach().numpy()[0] delta = self.r_last + self.gamma * Q_next - Q_cur Q_diff = self.Q(self.s_last, self.a_last) else: update = False # Update the parameter via the semi-gradient method if update: self.Q.zero_grad() Q_diff.backward() for p in self.Q.parameters(): p.data.add_(self.alpha * delta, p.grad.data) if not done: self.s_last = np.copy(s) self.a_last = np.copy(a) self.r_last = np.copy(r) # This is the environment env = SwingUpEnv() # For simplicity, we only consider forces of -1 and 1 numActions = 2 Actions = np.linspace(-1, 1, numActions) # This is our learning agent gamma = .95 agent = sarsaAgent(5, numActions, 20, 1, epsilon = 5e-2, gamma = gamma, alpha = 1e-5) maxSteps = 2e5 # This is a helper to deal with the fact that x[2] is actually an angle x_to_y = lambda x : np.array([x[0], x[1], np.cos(x[2]), np.sin(x[2]), x[3]]) R = [] UpTime = [] step = 0 ep = 0 while step < maxSteps: ep += 1 x = env.reset() C = 0. done = False t = 1 while not done: t += 1 step += 1 y = x_to_y(x) a = agent.action(y) u = Actions[a:a+1] x_next, c, done, info = env.step(u) max_up_time = info['max_up_time'] y_next = x_to_y(x_next) C += (1./t) * (c - C) agent.update(y, a, c, y_next, done) x = x_next if done: break if step >= maxSteps: break R.append(C) UpTime.append(max_up_time) #print('t:',ep+1,', R:',C,', L:',t-1,', G:',G,', Q:', Q_est, 'U:', max_up_time) print('Episode:',ep, 'Total Steps:',step, ', Ave. Reward:',C, ', Episode Length:',t-1, 'Max Up-Time:',max_up_time) env.close()