Notes updated 30/10/18 -- Add Adaptive Noise scaling method and Cross-entropy method( Evolution Strategies), black_box optimization technique (hill climbing)

Author: unnat5

CartPole_smart_agent.py

@@ -0,0 +1,55 @@
+## This Agent was trained using hill climbing with Adaptive Noise Scaling which comes under POLICY BASED METHODS.
+## And purpose of this script is to show a smart agent!
+import pickle
+import gym
+import numpy as np
+
+with open('hill_climbing_weight.pickle','rb') as f:
+    weight = pickle.load(f)
+
+
+class Policy():
+    def __init__(self, s_size=4, a_size=2):
+        """
+        Here I'm intializing the self.w with trained weights
+        The basic purpose of this function is to randomly initalize the weights for the network and then we would
+        optimize these weights with noise scaling(adaptive).
+        
+        Shape: [state_dimension,action_dimension] and softmax activation function at output layer-- when action space is discrete
+               [state_dimension, 1 node] and no activation function when action space is not discrete.
+        """
+        #self.w = 1e-4*np.random.rand(s_size, a_size) ##weights for simple linear policy: state_space x action_space
+        self.w = weight
+    
+    def forward(self, state):
+        """
+        Here we multipy(vectorized) our state with weights and get corresponding output 
+        """
+        x = np.dot(state,self.w)
+        ## below is the implementation of softmax function!!
+        return np.exp(x)/sum(np.exp(x))
+    
+    def act(self,state):
+        """
+        This function decides whether we want our policy to be stochastic or determinstic policy.
+        """
+        probs = self.forward(state)
+        #action = np.random.choice(2,p=probs)
+        action = np.argmax(probs)
+        # option 1: stochastic policy
+        # option 2: stochastic policy
+        return action
+    
+    
+policy = Policy()
+
+env = gym.make('CartPole-v0')
+
+for i in range(3):
+    state = env.reset()
+    while True:
+        env.render()
+        action = policy.act(state)
+        state,reward,done,_=env.step(action)
+        if done:
+            break