finish exercise 2.5

exercise_2_5.py

@@ -2,6 +2,8 @@ import numpy as np
 import random
 import matplotlib
 import matplotlib.pyplot as plt
+from tqdm import tqdm
+from mpl_toolkits.axes_grid1.axes_divider import make_axes_area_auto_adjustable
 
 class KArmedBandit:
     def __init__(self,
@@ -45,7 +47,13 @@ class KArmedBandit:
         if step >= self.steps or step < 0:
             ValueError("Step {} out of bounds. Current steps: {}".format(step, self.steps))
 
-        return self.rewards[action][step]
+        if self.random_walk:
+            # Slightly move all the averages
+            moves = np.random.normal(loc=0.0, scale=0.01, size=self.arms)
+            self.average_rewards = np.array(self.average_rewards) + moves
+
+        return np.random.normal(loc=self.average_rewards[action], scale=1, size=1)
+        # return self.rewards[action][step]
 
     def get_max_reward_action(self):
         return (max(self.average_rewards), np.argmax(self.average_rewards))
@@ -82,10 +90,15 @@ class Average:
         self.count += 1
 
 class Plotter:
-    def __init__(self, karmedbandit: KArmedBandit, epsilongreedy: EpsilonGreedy, steps: int):
+    def __init__(self, 
+                karmedbandit: KArmedBandit, 
+                epsilongreedy: EpsilonGreedy, 
+                steps: int,
+                alpha: bool = True):
         self.karmedbandit = karmedbandit
         self.agent = epsilongreedy
         self.steps = steps
+        self.alpha = alpha
         self.reset()
 
     def reset(self):
@@ -111,8 +124,8 @@ class Plotter:
         self.optimal_action_fraction[step].update(self.optimal_action_counter / (step+1))
 
     def run(self, runs=1):
-        for i in range(runs):
-            self.karmedbandit.generate_rewards()
+        for i in tqdm(range(runs), desc="Runs"):
+            self.karmedbandit.generate_average_rewards()
             # optimal_action = self.karmedbandit.get_max_reward_action()[1]
             self.agent.reset()
             self.actions_counter = [0] * self.karmedbandit.arms
@@ -121,8 +134,11 @@ class Plotter:
                 optimal_action = self.karmedbandit.get_max_reward_action()[1]
                 action = self.agent.choose_action()
                 reward = self.karmedbandit.get_reward(action, step)
-                # self.agent.update(action, reward, 1/(self.actions_counter[action]+1))
-                self.agent.update(action, reward, 0.1)
+                
+                if self.alpha:
+                    self.agent.update(action, reward, 0.1)
+                else:
+                    self.agent.update(action, reward, 1/(self.actions_counter[action]+1))
                 self.average_rewards[step].update(reward)
                 self.count_action(action)
                 if action == optimal_action:
@@ -138,27 +154,38 @@ class Plotter:
         violin_data = self.karmedbandit.sample_rewards(steps=10000)
         bp = ax1.violinplot(violin_data, showmeans=True, showmedians=False,
                             showextrema=False)
-        ax1.set_ylim([-4,4])
+        ax1.set_ylabel("Reward distribution", rotation='vertical')
+        # ax1.set_xlabel("Actions", rotation='horizontal')
 
         ax2.plot(range(self.steps), [x.value for x in self.average_rewards])
 
-        ylabel = "Average reward. Max average reward: {}".format(self.karmedbandit.get_max_reward_action()[0])
+        # ylabel = "Average reward. Max average reward: {}".format(self.karmedbandit.get_max_reward_action()[0])
+        ylabel = "Average reward."
         ax2.set(ylabel=ylabel, title='')
+        ax2.set_ylabel(ylabel, rotation='vertical')
+        # ax2.set_xlabel("Steps", rotation='horizontal')
         
         ax3.plot(range(self.steps), [x.value for x in self.optimal_action_fraction])
         ax3.set_ylim([0,1])
+        ax3.set_ylabel("Optimal action", rotation='vertical')
+        # ax3.set_xlabel("Steps", rotation='horizontal')
 
         plt.xlabel("Steps")
+        fig.set_figheight(8)
+        fig.set_figwidth(8)
+        # make_axes_area_auto_adjustable(ax1)
+        # make_axes_area_auto_adjustable(ax2)
+        # make_axes_area_auto_adjustable(ax3)
         fig.savefig("exercise_2-5.png")
         plt.show()
 
 
 if __name__ == "__main__":
     arms = 10
-    steps = 1000
-    armedbandit = KArmedBandit(arms=arms, steps=steps, random_walk=False)
+    steps = 10000
+    armedbandit = KArmedBandit(arms=arms, steps=steps, random_walk=True)
     greedy = EpsilonGreedy(0.1, arms)
 
     plotter = Plotter(armedbandit, greedy, steps)
-    plotter.run(runs=100)
+    plotter.run(runs=1000)
     plotter.plot()

exercises.md

@@ -142,4 +142,5 @@ Q_{n+1} = & \; Q_n + \alpha_n \left[R_n - Q_n\right] \\
 
 # Exercise 2.5
 See [exercise_2-5.py](./exercise_2-5.py) for code.
+![](./exercise_2-5.png)