[Reinforcement Learning] Policy Gradient (CartPole) 포스팅 즉
https://medium.com/@awjuliani/super-simple-reinforcement-learning-tutorial-part-2-ded33892c724
을 tensorflow 을 사용하지 않고 python 의 numpy 를 이용해 코딩해봤습니다.
python 3.6
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 | import numpy as np import gym class NN: def __init__(self): self.numHiddenLayerNeurons = 10 self.learningRate = 1e-2 self.discountFactorForReward = 0.99 self.inputDimension = 4 self.W1 = self.HeInitialization(self.inputDimension, self.numHiddenLayerNeurons) self.W2 = self.HeInitialization(self.numHiddenLayerNeurons, 1) self.W1GradientBuffer = np.zeros_like(self.W1) self.W2GradientBuffer = np.zeros_like(self.W2) def sigmoid(self,x): return 1.0 / (1.0 + np.exp(-x)) def dsigmoid(self,x): return x * (1. - x) def tanh(self,x): return np.tanh(x) def dtanh(self,x): return 1.0 - x * x def ReLU(self, x): return x * (x > 0) def dReLU(self,x): return 1.0 * (x > 0) def softmax(self, x): if x.ndim == 1: x = x.reshape([1, x.size]) modifiedX = x - np.max(x, 1).reshape([x.shape[0], 1]); sigmoid = np.exp(modifiedX) return sigmoid / np.sum(sigmoid, axis=1).reshape([sigmoid.shape[0], 1]); def XavierInitialization(self, NumIn, NumOut): return np.random.randn(NumIn, NumOut) / np.sqrt(NumIn) def HeInitialization(self, NumIn, NumOut): return np.random.randn(NumIn, NumOut) / np.sqrt(NumIn / 2) def feedForward(self, x): y1 = self.ReLU(np.matmul(x, self.W1)) score = np.matmul(y1, self.W2) probability = self.sigmoid(score) return y1, probability def backpropagation(self, x, error, y1, reward): discountedReward = self.discountReward(reward) discountedReward -= np.mean(discountedReward) discountedReward /= np.std(discountedReward) error *= discountedReward # dY2 = np.matmul(error, self.weights['W2'].T) dY2 = np.outer(error, self.W2) dY1 = self.dReLU(y1) dW1 = np.matmul(x.T, (dY2 * dY1)) dW2 = np.matmul(y1.T, error) self.W1GradientBuffer += dW1; self.W2GradientBuffer += dW2; def update(self): self.W1 += self.learningRate * self.W1GradientBuffer self.W2 += self.learningRate * self.W2GradientBuffer self.W1GradientBuffer = np.zeros_like(self.W1) self.W2GradientBuffer = np.zeros_like(self.W2) def discountReward(self, r): discounted_r = np.zeros_like(r) running_add = 0 for t in reversed(range(0, r.size)): running_add = running_add * self.discountFactorForReward + r[t] discounted_r[t] = running_add return discounted_r if __name__ == '__main__': batchSize = 5 env = gym.make('CartPole-v0') observation = env.reset() NN = NN() arrX, arrReward, arrY1, arrError = [], [], [], [] rewardSum = 0 episodeIndex = 1 env.reset() while episodeIndex <= 10000: x = np.reshape(observation, [1, NN.inputDimension]) y1, probability = NN.feedForward(x) action = 1 if np.random.uniform() < probability else 0 #e-greedy 필요할듯 arrX.append(x) arrY1.append(y1) arrError.append(action - probability) observation, reward, done, info = env.step(action) rewardSum += reward arrReward.append(reward) if done: episodeIndex += 1 episodeX = np.vstack(arrX) episodeReward = np.vstack(arrReward) episodeY1 = np.vstack(arrY1) episodeError = np.vstack(arrError) arrX, arrReward, arrY1, arrError = [], [], [], [] NN.backpropagation(episodeX, episodeError, episodeY1, episodeReward) if episodeIndex % batchSize == 0: NN.update() print('Average reward for episode %f. Total average reward %f.' % ( rewardSum / batchSize, rewardSum / batchSize)) if rewardSum / batchSize >= 200: print("Task solved in", episodeIndex, 'episodes!') break rewardSum = 0 observation = env.reset() | cs |