commit stale changes

.gitignore

@@ -0,0 +1 @@
+.idea/

k-bandits.py

@@ -113,7 +113,7 @@ def batch(k_arms, num_episode, num_problems, _param: Param):
 if __name__ == '__main__':
 	q_star = np.random.normal(0, 1, (num_realisations, k_arms))
 	arms = np.zeros((num_realisations, k_arms))
-	for i in range(k_arms):
+	for i in range(k_arms):pyt
 		arms[:,i] = np.random.normal(q_star[0, i], 1, num_realisations)  # first problem as a sample
 
 	pl.figure(figsize=(12, 8))
@@ -126,7 +126,7 @@ if __name__ == '__main__':
 	pl.figure(figsize=(12, 8))
 #	params = [Param(epsilon=0.0), Param(epsilon=0.01), Param(epsilon=0.1)]
 #	params = [Param(epsilon=0.1), Param(alpha=0.2, q_ic=5)]
-	params = [Param(epsilon=0.1), Param(argmax_func=ucb)]
+	params = [Param(epsilon=0.1), Param(argmax_func=ucb), Param(alpha=0.2, q_ic=5), Param(alpha=0.2, q_ic=5, argmax_func=ucb)]
 	legend = []
 	for param in params:
 		res_r, res_a = batch(k_arms, episode_len, num_realisations, param)

k-bandits_kaggle.py

@@ -0,0 +1,87 @@
+import numpy as np
+import matplotlib.pyplot as plt
+from tqdm import tqdm
+
+# Alternate implementation from:
+# https://www.kaggle.com/code/parsasam/reinforcement-learning-notes-multi-armed-bandits
+
+float_formatter = "{:.3f}".format
+np.set_printoptions(formatter={'float_kind': float_formatter})
+
+k = 10
+num_problems = 200
+num_steps = 2000
+
+q_star = np.random.normal(0, 1, (num_problems, k))
+arms = [0] * k
+
+for i in range(10):
+	arms[i] = np.random.normal(q_star[0, i], 1, 2000)  # first problem as a sample
+
+plt.figure(figsize=(12, 8))
+plt.ylabel('Rewards distribution')
+plt.xlabel('Actions')
+plt.xticks(range(1, 11))
+plt.yticks(np.arange(-5, 5, 0.5))
+
+plt.violinplot(arms, positions=range(1, 11), showmedians=True)
+plt.show()
+
+TIE_BREAK = 0.01
+
+def bandit(action, problem):
+	return np.random.normal(q_star[problem, action], 1)
+
+
+def simple_max(Q, N, t):
+	am = np.argmax(Q)
+	fm = Q == Q.max()
+	ffm = np.flatnonzero(fm)
+	return np.random.choice(ffm)  # breaking ties randomly
+
+
+def simple_bandit(k, epsilon, steps, initial_Q, alpha=0, argmax_func=simple_max):
+	rewards = np.zeros(steps)
+	actions = np.zeros(steps)
+
+	for i in tqdm(range(num_problems)):
+		Q = np.ones(k) * initial_Q  # initial Q
+		N = np.zeros(k)  # initalize number of rewards given
+		best_action = np.argmax(q_star[i])
+		for t in range(steps):
+			if np.random.rand() < epsilon:  # explore
+				a = np.random.randint(k)
+			else:  # exploit
+				a = argmax_func(Q, N, t)
+
+			reward = bandit(a, i)
+
+			N[a] += 1
+			if alpha > 0:
+				Q[a] = Q[a] + (reward - Q[a]) * alpha
+			else:
+				Q[a] = Q[a] + (reward - Q[a]) / N[a]
+
+			rewards[t] += reward
+
+			if a == best_action:
+				actions[t] += 1
+
+	return np.divide(rewards, num_problems), np.divide(actions, num_problems)
+
+
+def main():
+	ep_0, ac_0 = simple_bandit(k=10, epsilon=0, steps=num_steps, initial_Q=0)
+	ep_01, ac_01 = simple_bandit(k=10, epsilon=0.01, steps=num_steps, initial_Q=0)
+	ep_1, ac_1 = simple_bandit(k=10, epsilon=0.1, steps=num_steps, initial_Q=0)
+
+	plt.figure(figsize=(12,6))
+	plt.plot(ep_0, 'g', label='epsilon = 0')
+	plt.plot(ep_01, 'r', label='epsilon = 0.01')
+	plt.plot(ep_1, 'b', label='epsilon = 0.1')
+	plt.legend()
+	plt.show()
+
+
+if __name__ == '__main__':
+	main()