refactor solutions

.gitignore

@@ -2,13 +2,16 @@ _build/
 .DS_Store
 .ipynb_checkpoints/
 
+# exercise solutions
+solutions/
+
 # autogenerated Jupyter Book config for debugging
 conf.py
 
 # use MyST markdown as authoritative source
 *.ipynb
 
-# latex
+# latex build outputs
 *.aux
 *.bbl
 *.bcf

Makefile

@@ -16,17 +16,19 @@ _META = \
 
 META = $(addsuffix .md, $(addprefix book/, $(_META)))
 
-CHAPTERS = $(NOTEBOOKS) $(META)
+SOLUTIONS = book/solutions/bandits.py
+
+SOURCE = $(NOTEBOOKS) $(META) $(SOLUTIONS)
 
 CONFIG = book/_config.yml book/_toc.yml
 
-book/_build/html: $(CHAPTERS) $(CONFIG)
+book/_build/html: $(SOURCE) $(CONFIG)
 	$(RUN) jb build book
 
 open: book/_build/html
 	open book/_build/html/index.html
 
-book/_build/latex: $(CHAPTERS) $(CONFIG)
+book/_build/latex: $(SOURCE) $(CONFIG)
 	$(RUN) jb build book --builder latex
 
 pdf: book/_build/latex

book/bandits.md

@@ -26,6 +26,8 @@ from abc import ABC, abstractmethod  # "Abstract Base Class"
 from typing import Callable, Union
 import matplotlib.pyplot as plt
 
+import solutions.bandits as solutions
+
 np.random.seed(184)
 
 # output_notebook()  # set up bokeh
@@ -38,14 +40,6 @@ def random_argmax(ary: Array) -> int:
     return np.random.choice(max_idx).item()
 
 
-def choose_zero(ary: Float[Array, " K"]) -> Union[int, None]:
-    min_idx = np.flatnonzero(ary == 0)
-    if min_idx.size > 0:
-        return np.random.choice(min_idx).item()
-    else:
-        return None
-
-
 latex = latexify.algorithmic(
     prefixes={"mab"},
     identifiers={"arm": "a_t", "reward": "r", "means": "mu"},
@@ -205,6 +199,8 @@ The rest of the chapter comprises a series of increasingly sophisticated
 MAB algorithms.
 
 ```{code-cell}
+:tags: [hide-input]
+
 def plot_strategy(mab: MAB, agent: Agent):
     plt.figure(figsize=(10, 6))
 
@@ -236,7 +232,7 @@ exploration").
 class PureExploration(Agent):
     def choose_arm(self):
         """Choose an arm uniformly at random."""
-        return np.random.randint(self.K)
+        return solutions.pure_exploration_choose_arm(self)
 ```
 
 Note that
@@ -274,15 +270,7 @@ call this the **pure greedy** strategy.
 class PureGreedy(Agent):
     def choose_arm(self):
         """Choose the arm with the highest observed reward on its first pull."""
-        if self.count < self.K:
-            # first K steps: choose each arm once
-            return self.count
-
-        if self.count == self.K:
-            # after the first K steps, choose the arm with the highest observed reward
-            self.greedy_arm = random_argmax(self.history[:, 1])
-
-        return self.greedy_arm
+        return solutions.pure_greedy_choose_arm(self)
 ```
 
 Note we’ve used superscripts $r^k$ during the exploration phase to
@@ -332,15 +320,7 @@ class ExploreThenCommit(Agent):
         self.N_explore = N_explore
 
     def choose_arm(self):
-        if self.count < self.K * self.N_explore:
-            # exploration phase: choose each arm N_explore times
-            return self.count // self.N_explore
-
-        # exploitation phase: choose the arm with the highest observed reward
-        if self.count == self.K * self.N_explore:
-            self.greedy_arm = random_argmax(self.history[:, 1])
-
-        return self.greedy_arm
+        return solutions.etc_choose_arm(self)
 ```
 
 ```{code-cell}
@@ -493,16 +473,7 @@ class EpsilonGreedy(Agent):
         self.get_epsilon = get_epsilon
 
     def choose_arm(self):
-        epsilon = self.get_epsilon(self.count)
-        if np.random.random() < epsilon:
-            return np.random.randint(0, self.K - 1)
-        else:
-            counts = self.history.sum(axis=1)
-            unvisited = choose_zero(counts)
-            if unvisited is not None:
-                return unvisited
-            sample_means = self.history[:, 1] / counts
-            return random_argmax(sample_means)
+        return solutions.epsilon_greedy_choose_arm(self)
 ```
 
 ```{code-cell}
@@ -600,14 +571,7 @@ class UCB(Agent):
         self.delta = delta
 
     def choose_arm(self):
-        counts = self.history.sum(axis=1)
-        unvisited = choose_zero(counts)
-        if unvisited is not None:
-            return unvisited
-        sample_means = self.history[:, 1] / counts
-        bounds = np.sqrt(np.log(2 * self.count / self.delta) / (2 * counts))
-        ucbs = sample_means + bounds
-        return random_argmax(ucbs)
+        return solutions.ucb_choose_arm(self)
 ```
 
 Intuitively, UCB prioritizes arms where: