This website works better with desktop in both themes, for mobile devices please change to light theme.

Multi-Armed Bandit (k-Armed Bandit)#

References#

[1]:
from datetime import timedelta, datetime
import time

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns

from IPython.display import display, clear_output

Introduction#

  • Iteratively selects one of the multiple fixed choices, These choices are called arms or actions or levers

  • Property of each choice is partially known at the time of allocation but become better understood as the time passes.

    • Either we know it earlier (Partially)

    • or we’ll know it with time

  • these arms are Independent

  • This algorithm works as Best Arm Identification

  • Target here is to Minimize Regret

  • Exploration \(\leftrightarrow\) Exploitation

Objective > Maximize the sum of rewards earned through a sequence of arm pulls. > > Each machines provides a random reward from a probability distribution specific to the machine(not known apriori)

Reward Models#

Reward

Description

IID(Independent & Identically Distributed) Rewards

Reward from each arm is drawn independently from a fixed distribution that depends on the arm and not the round t

Adverserial Rewards

Rewards can be arbitrary as they are chosen by an adversary trying to fool the algorithm, May be oblivious or adaptive

Constrained Adversary

Rewards are chosen from an adversary that is subjected to some constraint

Random Process Rewards

An arm’s state, which determines reward, evolves over time as a random process state transition

Feedbacks#

Feedback

Description

Bandit Feedback

algorithm observes the reward only for chosen arm

Full Feedback

algorithm observes the reward for all the arms that could have been chosen

Partial Feedback

some info is revealed + reward from the chosen arm (doesn’t always amount to full feedback)

Contexts#

  • Algorithm may observe context before chosing an arm

  • Personalized actions

News Website : user location & demographics

Dynamic Pricing : customer’s device, location, demographics

Bayesian Priors#

Problem instance comes from known distribution called Bayesian Prior

Stochastic Bandits#

  • IID Rewards

  • \(K\) Possible Arms/ Actions

  • \(T\) Rounds

  • Maximize total Reward over T rounds

  • Algorithm observes reward only from selected arms Bandit Feedback

Arms : \(K\)#

  • Possible actions/arms

  • Follows IID

  • Chosen over T rounds

[2]:
class Arms:
    @staticmethod
    def arm_1(mu=10, sigma=5):
        return np.random.normal(mu, sigma)

    @staticmethod
    def arm_2(mu=8, sigma=3):
        return np.random.normal(mu, sigma)

    @staticmethod
    def arm_3(mu=5, sigma=25):
        return np.random.normal(mu, sigma)


arms = ["arm_1", "arm_2", "arm_3"]
[3]:
def arms_cyclic_iter(arms):
    i = 0
    while True:
        if i >= len(arms):
            i = 0
        yield arms[i]
        i += 1

def print_df(df, sleep_time):
    clear_output(wait=True)
    display(df)
    time.sleep(sleep_time)
[4]:
def random_choose_arm(arms, size=None, replace=True, p=None):
    return np.random.choice(arms, size=size, replace=replace, p=p)

Reward Distribution : \(D_a\)#

  • Reward for each action is IID, for each action \(a\) there is a distribution \(D_a\) called Reward Distribution

  • Each time the action is chosen and the reward is sampled from a dsitribution that is Initially Unknown

  • \(K\) arms, \(T\) rounds (Known), \(D_a\) for arm \(a\) (unknown)

  • at \(t \in [T]\)

    • pick arm \(a_t\)

    • reward \(r_t \in [0, 1]\) sampled from \(D_a, a=a_t\)

    • collects reward \(r_t\) & observes nothing else

  • Mean reward value \(\mu \in [0, 1]\)

  • \(\mu(a) = \mathbb{E}[D_a]\) (Mean of reward from action = expected value for reward distribution)

  • Example Bernoulli Distribution

  • below example of reward distribution is not bound in [0, 1] but shows different arms with different rewards distribution

[5]:
fig, ax = plt.subplots(1, 1, figsize=(8, 3))

params = {
    "kde": True,
    "element": "poly",
    "ax": ax,
}
sns.histplot([Arms.arm_1() for _ in range(10000)], label="arm 1", **params)
sns.histplot([Arms.arm_2() for _ in range(10000)], label="arm 2", **params)
sns.histplot([Arms.arm_3() for _ in range(10000)], label="arm 3", **params)

fig.suptitle("Reward Distribution")
fig.legend()
fig.show()
../_images/RL_mab_11_0.png

Best Mean Reward/ Optimum Reward : \(\mu^*\)#

\begin{align} \mu(a) &= \mathbb{E}[D_a] \\ \text{Optimum Reward } \mu^* &= {\text{max}\atop{a \in \mathbb{A}}} \mu(a) \end{align}

Gap in Arm : \(\Delta(a)\)#

gap in arm a is defined as difference of best mean reward and reward of action a \begin{align} \Delta(a) &:= \mu^* - \mu(a) \end{align}

Best Arm/ Optimal Arm : \(a^*\)#

optimal arm is arm with best mean reward

\begin{align} a^* &:= \mu(a) = \mu^* \rightarrow \Delta(a) \approx 0 \end{align}

[6]:
def optimal_arm(df, best_mean_reward):
    return df[df["reward_mean"] == best_mean_reward]["arm"].iloc[0]

Regret : \(R(T)\)#

  • Compare cumulative reward (Mean reward at t step cumulative sum till round T) to best arm benchmark (Best Mean Reward T Rounds times), what if we play same best arm for all rounds \(T\). (Always playing optimal arm) \begin{align} \text{Regret at round T} \\ R(T) = \mu^* . T - \Sigma_{t=1}^{T} \mu(a_t) \\ \end{align}

[7]:
def calculate_regret(cumulative_reward, best_mean_reward, T):
    best_arm_benchmark = best_mean_reward * T
    regret = best_arm_benchmark - cumulative_reward
    return regret
[8]:
def get_metrics(rewardsdf: pd.DataFrame, verbose=True):
    aggdf = (
        rewardsdf.groupby(by=["arm", "decision"], as_index=False)
        .agg(
            ts_count=("ts", "count"),
            ts_min=("ts", "min"),
            reward_min=("reward", "min"),
            reward_mean=("reward", "mean"),
            reward_max=("reward", "max"),
            reward_sum=("reward", "sum"),
            reward_std=("reward", "std"),
        )
        .sort_values(by=["ts_min"])
    )

    best_mean_reward = aggdf["reward_mean"].max()
    cumulative_reward = aggdf["reward_sum"].sum()
    T = aggdf["ts_count"].sum()

    optimum_reward = best_mean_reward * T

    best_arm = optimal_arm(aggdf, best_mean_reward)

    regret = calculate_regret(
        cumulative_reward=cumulative_reward, best_mean_reward=best_mean_reward, T=T
    )
    if verbose:
        print_df(aggdf, 0.2)
        print(
            f"""
        Rounds              : {T}
        Optimum Reward      : {optimum_reward}
        Cumulative Reward   : {cumulative_reward}
        Regret              : {regret}
        """
        )
    return (best_arm, aggdf, optimum_reward, cumulative_reward, regret)
[9]:
def plot_reward(rewarddf, x="ts"):
    rewarddf["decision-arm"] = rewarddf["decision"] + "-" + rewarddf["arm"]
    fig, ax = plt.subplots(2, 1, figsize=(15, 8))

    line_plot_params = {
        "alpha": 0.6,
        "color": "grey",
        "linewidth": 0.5,
    }

    sns.lineplot(data=rewarddf, x=x, y="reward", ax=ax[0], **line_plot_params)
    sns.scatterplot(data=rewarddf, x=x, y="reward", hue="decision-arm", ax=ax[0])

    sns.lineplot(data=rewarddf, x=x, y="reward_cumsum", ax=ax[1], **line_plot_params)
    sns.scatterplot(
        data=rewarddf, x=x, y="reward_cumsum", hue="decision-arm", ax=ax[1]
    )

    fig.autofmt_xdate(rotation=45)
    fig.tight_layout()
    return fig

Model#

Code below is a rudimentary implementation of how a MAB model will collect data and fit distribution.

[10]:
class Model:
    def __init__(self, arms) -> None:
        self.ts_idx = datetime(2024, 1, 1, 0, 0, 0)
        self.timedelta = timedelta(minutes=30)
        self.seq_idx = 0
        self.decision_idx = 0
        self.rewards = pd.DataFrame(
            columns=["decision_idx", "decision", "seq", "ts", "arm", "reward"]
        )
        self.arms = arms
        self.chosen_arm = None

    def get_decision_idx(self, decision):

        if (
            len(self.rewards) == 0
            or self.rewards["decision"].iloc[len(self.rewards) - 1] != decision
        ):
            self.decision_idx += 1
        return self.decision_idx

    def get_arm(self):
        if self.chosen_arm is None:
            if len(self.rewards) == 0:
                raise ValueError("First run explore to get some data")
            self.chosen_arm = get_metrics(self.rewards, verbose=False)[0]
        return self.chosen_arm

    def update_rewards(self, reward_payload):
        self.rewards.loc[len(self.rewards)] = reward_payload
        print_df(self.rewardsdf, 0.2)

    def explore(self, enforce_arm=None):
        self.chosen_arm = None
        explore_arm = enforce_arm or random_choose_arm(self.arms)
        reward = getattr(Arms, explore_arm)()

        self.ts_idx += self.timedelta
        self.seq_idx += 1

        decision = "explore"
        decision_idx = self.get_decision_idx(decision)
        self.update_rewards(
            [decision_idx, decision, self.seq_idx, self.ts_idx, explore_arm, reward]
        )

    def exploit(self):
        chosen_arm = self.get_arm()
        reward = getattr(Arms, chosen_arm)()

        self.ts_idx += self.timedelta
        self.seq_idx += 1

        decision = "exploit"
        decision_idx = self.get_decision_idx(decision)
        self.update_rewards(
            [decision_idx, decision, self.seq_idx, self.ts_idx, chosen_arm, reward]
        )

    @property
    def rewardsdf(self):
        df = self.rewards.copy()
        df.insert(len(df.columns), column="reward_cumsum", value=df["reward"].cumsum())
        return df

Only Exploration#

Algorithm#

  1. \(K\) arms, \(T\) rounds

  2. In each round \(t \in T\)

  1. Algorithm picks an arm \(a_t\) (Stochastic/ Sequential)

  2. Reward \(r_t\) is sampled/observed from distribution \(D_a\) where \(a=a_T\)

  3. Algorithm collects reward \(r_t\), observes nothing else(Bandit feedback)

Stochastic#

[11]:
model = Model(arms)

for _ in range(100):
    model.explore()
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_3 -11.788321 -11.788321
1 1 explore 2 2024-01-01 01:00:00 arm_2 12.088685 0.300363
2 1 explore 3 2024-01-01 01:30:00 arm_1 9.827448 10.127812
3 1 explore 4 2024-01-01 02:00:00 arm_2 10.922434 21.050246
4 1 explore 5 2024-01-01 02:30:00 arm_1 8.885288 29.935534
... ... ... ... ... ... ... ...
95 1 explore 96 2024-01-03 00:00:00 arm_3 -1.721018 854.314202
96 1 explore 97 2024-01-03 00:30:00 arm_3 6.783670 861.097873
97 1 explore 98 2024-01-03 01:00:00 arm_3 -39.009226 822.088647
98 1 explore 99 2024-01-03 01:30:00 arm_2 10.385623 832.474270
99 1 explore 100 2024-01-03 02:00:00 arm_1 14.332681 846.806951

100 rows × 7 columns

[12]:
explore_rewarddf = model.rewardsdf
get_metrics(explore_rewarddf)
plot_reward(explore_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
2 arm_3 explore 40 2024-01-01 00:30:00 -40.480190 8.119206 46.224011 324.768249 23.625556
1 arm_2 explore 37 2024-01-01 01:00:00 1.412995 8.095527 13.555356 299.534482 3.103868
0 arm_1 explore 23 2024-01-01 01:30:00 3.240814 9.674097 18.647836 222.504220 4.188599

        Rounds              : 100
        Optimum Reward      : 967.4096520112096
        Cumulative Reward   : 846.806950843292
        Regret              : 120.60270116791753

../_images/RL_mab_25_2.png

Sequential#

[13]:
arms_iterator = arms_cyclic_iter(arms)
model = Model(arms)

for _ in range(100):
    model.explore(next(arms_iterator))
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 14.717441 14.717441
1 1 explore 2 2024-01-01 01:00:00 arm_2 10.151536 24.868977
2 1 explore 3 2024-01-01 01:30:00 arm_3 34.513253 59.382230
3 1 explore 4 2024-01-01 02:00:00 arm_1 7.857376 67.239606
4 1 explore 5 2024-01-01 02:30:00 arm_2 9.939486 77.179092
... ... ... ... ... ... ... ...
95 1 explore 96 2024-01-03 00:00:00 arm_3 -0.086446 699.868702
96 1 explore 97 2024-01-03 00:30:00 arm_1 11.695269 711.563971
97 1 explore 98 2024-01-03 01:00:00 arm_2 3.749240 715.313211
98 1 explore 99 2024-01-03 01:30:00 arm_3 -5.496441 709.816770
99 1 explore 100 2024-01-03 02:00:00 arm_1 16.264725 726.081496

100 rows × 7 columns

[14]:
explore_rewarddf = model.rewardsdf
get_metrics(explore_rewarddf)
plot_reward(explore_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
0 arm_1 explore 34 2024-01-01 00:30:00 -0.531654 10.507842 19.903238 357.266638 4.991634
1 arm_2 explore 33 2024-01-01 01:00:00 0.727823 8.089693 14.239782 266.959881 3.288865
2 arm_3 explore 33 2024-01-01 01:30:00 -47.006049 3.086514 54.848921 101.854976 25.057924

        Rounds              : 100
        Optimum Reward      : 1050.784229925647
        Cumulative Reward   : 726.0814956762381
        Regret              : 324.7027342494089

../_images/RL_mab_28_2.png

Uniform Exploration#

Algorithm#

  1. For \(K\) arms, \(T\) rounds

  2. Exploration Phase : Try each arm N times

  3. Select an arm \(\hat{a}\) with highest mean reward

  4. Exploitation Phase : use \(\hat{a}\) for rest of the rounds

[15]:
arms_iterator = arms_cyclic_iter(arms)

model = Model(arms)

for _ in range(3):
    model.explore(next(arms_iterator))

for i in range(97):
    model.exploit()
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 5.879398 5.879398
1 1 explore 2 2024-01-01 01:00:00 arm_2 8.022709 13.902108
2 1 explore 3 2024-01-01 01:30:00 arm_3 4.994972 18.897080
3 2 exploit 4 2024-01-01 02:00:00 arm_2 6.281499 25.178578
4 2 exploit 5 2024-01-01 02:30:00 arm_2 8.768576 33.947154
... ... ... ... ... ... ... ...
95 2 exploit 96 2024-01-03 00:00:00 arm_2 5.323829 806.060974
96 2 exploit 97 2024-01-03 00:30:00 arm_2 10.659535 816.720509
97 2 exploit 98 2024-01-03 01:00:00 arm_2 5.718620 822.439129
98 2 exploit 99 2024-01-03 01:30:00 arm_2 6.567910 829.007039
99 2 exploit 100 2024-01-03 02:00:00 arm_2 9.989318 838.996357

100 rows × 7 columns

[16]:
exploit_rewarddf = model.rewardsdf
get_metrics(exploit_rewarddf)
plot_reward(exploit_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
0 arm_1 explore 1 2024-01-01 00:30:00 5.879398 5.879398 5.879398 5.879398 NaN
2 arm_2 explore 1 2024-01-01 01:00:00 8.022709 8.022709 8.022709 8.022709 NaN
3 arm_3 explore 1 2024-01-01 01:30:00 4.994972 4.994972 4.994972 4.994972 NaN
1 arm_2 exploit 97 2024-01-01 02:00:00 -1.311294 8.454632 16.735225 820.099277 3.371759

        Rounds              : 100
        Optimum Reward      : 845.4631722542013
        Cumulative Reward   : 838.9963569129461
        Regret              : 6.466815341255142

../_images/RL_mab_31_2.png

Selective Exploitation#

Arm 1#

[17]:
model = Model(arms)

for _ in range(3):
    model.explore(next(arms_iterator))

model.chosen_arm = "arm_1"

for i in range(97):
    model.exploit()
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 1.584226 1.584226
1 1 explore 2 2024-01-01 01:00:00 arm_2 4.575686 6.159912
2 1 explore 3 2024-01-01 01:30:00 arm_3 -46.110331 -39.950419
3 2 exploit 4 2024-01-01 02:00:00 arm_1 8.441025 -31.509394
4 2 exploit 5 2024-01-01 02:30:00 arm_1 16.706485 -14.802909
... ... ... ... ... ... ... ...
95 2 exploit 96 2024-01-03 00:00:00 arm_1 -0.027129 887.467501
96 2 exploit 97 2024-01-03 00:30:00 arm_1 8.294347 895.761848
97 2 exploit 98 2024-01-03 01:00:00 arm_1 13.948062 909.709910
98 2 exploit 99 2024-01-03 01:30:00 arm_1 8.986954 918.696864
99 2 exploit 100 2024-01-03 02:00:00 arm_1 12.077487 930.774351

100 rows × 7 columns

[18]:
exploit_rewarddf = model.rewardsdf
get_metrics(exploit_rewarddf)
plot_reward(exploit_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
1 arm_1 explore 1 2024-01-01 00:30:00 1.584226 1.584226 1.584226 1.584226 NaN
2 arm_2 explore 1 2024-01-01 01:00:00 4.575686 4.575686 4.575686 4.575686 NaN
3 arm_3 explore 1 2024-01-01 01:30:00 -46.110331 -46.110331 -46.110331 -46.110331 NaN
0 arm_1 exploit 97 2024-01-01 02:00:00 -0.276465 10.007472 23.195923 970.724770 5.213823

        Rounds              : 100
        Optimum Reward      : 1000.7471851949623
        Cumulative Reward   : 930.7743505778576
        Regret              : 69.9728346171047

../_images/RL_mab_35_2.png

Arm 2#

[19]:
model = Model(arms)

for _ in range(3):
    model.explore(next(arms_iterator))

model.chosen_arm = "arm_2"

for i in range(97):
    model.exploit()
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 14.293605 14.293605
1 1 explore 2 2024-01-01 01:00:00 arm_2 5.017779 19.311384
2 1 explore 3 2024-01-01 01:30:00 arm_3 -16.783069 2.528316
3 2 exploit 4 2024-01-01 02:00:00 arm_2 4.583187 7.111502
4 2 exploit 5 2024-01-01 02:30:00 arm_2 5.678105 12.789607
... ... ... ... ... ... ... ...
95 2 exploit 96 2024-01-03 00:00:00 arm_2 8.846101 788.602028
96 2 exploit 97 2024-01-03 00:30:00 arm_2 3.746286 792.348314
97 2 exploit 98 2024-01-03 01:00:00 arm_2 2.557700 794.906013
98 2 exploit 99 2024-01-03 01:30:00 arm_2 5.727292 800.633305
99 2 exploit 100 2024-01-03 02:00:00 arm_2 9.785695 810.419000

100 rows × 7 columns

[20]:
exploit_rewarddf = model.rewardsdf
get_metrics(exploit_rewarddf)
plot_reward(exploit_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
0 arm_1 explore 1 2024-01-01 00:30:00 14.293605 14.293605 14.293605 14.293605 NaN
2 arm_2 explore 1 2024-01-01 01:00:00 5.017779 5.017779 5.017779 5.017779 NaN
3 arm_3 explore 1 2024-01-01 01:30:00 -16.783069 -16.783069 -16.783069 -16.783069 NaN
1 arm_2 exploit 97 2024-01-01 02:00:00 0.325208 8.328770 16.413429 807.890684 3.349533

        Rounds              : 100
        Optimum Reward      : 1429.3605254607367
        Cumulative Reward   : 810.4189997093333
        Regret              : 618.9415257514033

../_images/RL_mab_38_2.png

Arm 3#

[21]:
model = Model(arms)

for _ in range(3):
    model.explore(next(arms_iterator))

model.chosen_arm = "arm_3"

for i in range(97):
    model.exploit()
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 9.701172 9.701172
1 1 explore 2 2024-01-01 01:00:00 arm_2 8.269244 17.970416
2 1 explore 3 2024-01-01 01:30:00 arm_3 31.149943 49.120360
3 2 exploit 4 2024-01-01 02:00:00 arm_3 -54.441744 -5.321384
4 2 exploit 5 2024-01-01 02:30:00 arm_3 -24.978085 -30.299469
... ... ... ... ... ... ... ...
95 2 exploit 96 2024-01-03 00:00:00 arm_3 11.065099 835.878021
96 2 exploit 97 2024-01-03 00:30:00 arm_3 -15.780444 820.097577
97 2 exploit 98 2024-01-03 01:00:00 arm_3 5.832560 825.930137
98 2 exploit 99 2024-01-03 01:30:00 arm_3 -19.050207 806.879931
99 2 exploit 100 2024-01-03 02:00:00 arm_3 16.246434 823.126365

100 rows × 7 columns

[22]:
exploit_rewarddf = model.rewardsdf
get_metrics(exploit_rewarddf)
plot_reward(exploit_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
0 arm_1 explore 1 2024-01-01 00:30:00 9.701172 9.701172 9.701172 9.701172 NaN
1 arm_2 explore 1 2024-01-01 01:00:00 8.269244 8.269244 8.269244 8.269244 NaN
3 arm_3 explore 1 2024-01-01 01:30:00 31.149943 31.149943 31.149943 31.149943 NaN
2 arm_3 exploit 97 2024-01-01 02:00:00 -54.441744 7.979443 57.080189 774.006005 23.401965

        Rounds              : 100
        Optimum Reward      : 3114.9943150453537
        Cumulative Reward   : 823.1263646215909
        Regret              : 2291.8679504237625

../_images/RL_mab_41_2.png

Alternative Sequentially Exploration & Exploitation#

[23]:
def explore_exploit(
    n_iterations=100,
    explore_iters=10,
    exploit_iters=10,
    arms=arms,
):

    model = Model(arms)
    run_explore = True

    explore_counter, exploit_counter = 0, 0

    for i in range(n_iterations):
        if run_explore:
            model.explore()
            explore_counter += 1
        else:
            model.exploit()
            exploit_counter += 1

        if explore_counter == explore_iters:
            run_explore = False
            explore_counter = 0

        if exploit_counter == exploit_iters:
            run_explore = True
            exploit_counter = 0

    return model

Equally Explore & Exploit#

[24]:
model = explore_exploit(
    n_iterations=100,
    explore_iters=10,
    exploit_iters=10,
    arms=["arm_1", "arm_2", "arm_3"],
)
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_3 -39.184432 -39.184432
1 1 explore 2 2024-01-01 01:00:00 arm_3 -2.250016 -41.434448
2 1 explore 3 2024-01-01 01:30:00 arm_3 8.326022 -33.108426
3 1 explore 4 2024-01-01 02:00:00 arm_3 15.703394 -17.405032
4 1 explore 5 2024-01-01 02:30:00 arm_1 10.646341 -6.758691
... ... ... ... ... ... ... ...
95 10 exploit 96 2024-01-03 00:00:00 arm_1 9.047375 786.614477
96 10 exploit 97 2024-01-03 00:30:00 arm_1 3.454034 790.068511
97 10 exploit 98 2024-01-03 01:00:00 arm_1 7.935937 798.004448
98 10 exploit 99 2024-01-03 01:30:00 arm_1 5.877367 803.881815
99 10 exploit 100 2024-01-03 02:00:00 arm_1 18.039669 821.921484

100 rows × 7 columns

[25]:
experiment_rewarddf = model.rewardsdf
get_metrics(experiment_rewarddf)
plot_reward(experiment_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
3 arm_3 explore 20 2024-01-01 00:30:00 -62.780166 2.378776 47.756459 47.575522 24.297217
1 arm_1 explore 19 2024-01-01 02:30:00 1.944325 10.068205 18.025927 191.295897 4.819601
2 arm_2 explore 11 2024-01-01 03:00:00 2.590968 7.912387 14.106835 87.036258 4.005542
0 arm_1 exploit 50 2024-01-01 05:30:00 3.263505 9.920276 21.659107 496.013807 4.229813

        Rounds              : 100
        Optimum Reward      : 1006.8205081327129
        Cumulative Reward   : 821.9214841863763
        Regret              : 184.89902394633657

../_images/RL_mab_46_2.png

Less Explore & More Exploit#

[26]:
model = explore_exploit(
    n_iterations=100,
    explore_iters=5,
    exploit_iters=10,
    arms=arms,
)
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_2 9.242226 9.242226
1 1 explore 2 2024-01-01 01:00:00 arm_3 6.834996 16.077222
2 1 explore 3 2024-01-01 01:30:00 arm_2 3.984192 20.061414
3 1 explore 4 2024-01-01 02:00:00 arm_1 7.456886 27.518300
4 1 explore 5 2024-01-01 02:30:00 arm_1 3.323603 30.841903
... ... ... ... ... ... ... ...
95 14 exploit 96 2024-01-03 00:00:00 arm_2 2.434778 681.921263
96 14 exploit 97 2024-01-03 00:30:00 arm_2 6.428046 688.349309
97 14 exploit 98 2024-01-03 01:00:00 arm_2 8.146294 696.495603
98 14 exploit 99 2024-01-03 01:30:00 arm_2 6.032236 702.527839
99 14 exploit 100 2024-01-03 02:00:00 arm_2 10.112191 712.640030

100 rows × 7 columns

[27]:
experiment_rewarddf = model.rewardsdf
get_metrics(experiment_rewarddf)
plot_reward(experiment_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
2 arm_2 explore 16 2024-01-01 00:30:00 3.321153 9.472455 12.793415 151.559277 2.765254
4 arm_3 explore 11 2024-01-01 01:00:00 -11.956004 5.966866 38.700383 65.635525 15.065037
0 arm_1 explore 8 2024-01-01 02:00:00 2.946490 8.701732 19.989582 69.613853 5.662205
3 arm_3 exploit 10 2024-01-01 03:00:00 -34.573848 -1.453645 37.609636 -14.536448 24.937441
1 arm_2 exploit 55 2024-01-01 10:30:00 1.818452 8.006688 14.404394 440.367822 3.353643

        Rounds              : 100
        Optimum Reward      : 947.2454819980799
        Cumulative Reward   : 712.6400299643165
        Regret              : 234.6054520337634

../_images/RL_mab_49_2.png

Epsilon (\(\epsilon\)) Greedy#

Algorithm#

  1. for \(K\) arms, \(T\) rounds

  2. for each round \(t \in [T]\)

  1. Toss a coin and generate success probability \(\epsilon_t\)

  1. if success then explore : choose any arm uniformly at random

  2. if failure then exploit : choose an arm with highest average rewards so far

[28]:
def epsilon_greedy(n_iterations=100, epsilon=0.2, arms=arms):

    model = Model(arms)

    model.explore()

    for i in range(n_iterations):

        p = np.random.random()

        if p < epsilon:
            model.explore()
        else:
            model.exploit()

    return model

\(\epsilon\) = 0.1#

[29]:
model = epsilon_greedy(n_iterations=100, epsilon=0.1, arms=arms)
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_1 13.164255 13.164255
1 1 explore 2 2024-01-01 01:00:00 arm_1 7.513586 20.677841
2 2 exploit 3 2024-01-01 01:30:00 arm_1 14.617776 35.295617
3 2 exploit 4 2024-01-01 02:00:00 arm_1 5.952256 41.247873
4 2 exploit 5 2024-01-01 02:30:00 arm_1 3.362488 44.610361
... ... ... ... ... ... ... ...
96 18 exploit 97 2024-01-03 00:30:00 arm_1 13.497176 923.444317
97 18 exploit 98 2024-01-03 01:00:00 arm_1 9.582637 933.026954
98 18 exploit 99 2024-01-03 01:30:00 arm_1 1.226871 934.253825
99 18 exploit 100 2024-01-03 02:00:00 arm_1 25.016123 959.269948
100 18 exploit 101 2024-01-03 02:30:00 arm_1 9.758156 969.028104

101 rows × 7 columns

[30]:
experiment_rewarddf = model.rewardsdf
get_metrics(experiment_rewarddf)
plot_reward(experiment_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
1 arm_1 explore 5 2024-01-01 00:30:00 7.513586 11.000555 13.164255 55.002775 2.568003
0 arm_1 exploit 33 2024-01-01 01:30:00 1.226871 11.041362 25.016123 364.364938 4.833072
4 arm_3 explore 2 2024-01-01 07:30:00 -3.164411 10.910349 24.985109 21.820698 19.904717
3 arm_3 exploit 58 2024-01-01 08:00:00 -38.117554 8.676521 56.969118 503.238221 24.102827
2 arm_2 explore 3 2024-01-01 10:30:00 8.029772 8.200490 8.305829 24.601471 0.149190

        Rounds              : 101
        Optimum Reward      : 1115.1775375160755
        Cumulative Reward   : 969.0281035361436
        Regret              : 146.14943397993193

../_images/RL_mab_54_2.png

\(\epsilon\) = 0.2#

[31]:
model = epsilon_greedy(n_iterations=100, epsilon=0.2, arms=arms)
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_2 9.872633 9.872633
1 2 exploit 2 2024-01-01 01:00:00 arm_2 5.260020 15.132653
2 2 exploit 3 2024-01-01 01:30:00 arm_2 10.135395 25.268048
3 2 exploit 4 2024-01-01 02:00:00 arm_2 8.162236 33.430284
4 2 exploit 5 2024-01-01 02:30:00 arm_2 8.573806 42.004090
... ... ... ... ... ... ... ...
96 29 explore 97 2024-01-03 00:30:00 arm_1 9.712263 655.099951
97 30 exploit 98 2024-01-03 01:00:00 arm_3 -1.338992 653.760959
98 30 exploit 99 2024-01-03 01:30:00 arm_3 -4.951924 648.809035
99 30 exploit 100 2024-01-03 02:00:00 arm_3 40.488164 689.297199
100 30 exploit 101 2024-01-03 02:30:00 arm_3 33.910435 723.207635

101 rows × 7 columns

[32]:
experiment_rewarddf = model.rewardsdf
get_metrics(experiment_rewarddf)
plot_reward(experiment_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
2 arm_2 explore 10 2024-01-01 00:30:00 3.743866 8.093269 13.298554 80.932688 2.887230
1 arm_2 exploit 8 2024-01-01 01:00:00 4.105006 7.720173 12.241799 61.761386 2.673941
4 arm_3 explore 3 2024-01-01 05:00:00 12.307630 21.670430 31.019323 65.011289 9.355854
3 arm_3 exploit 77 2024-01-01 05:30:00 -44.760957 6.359283 54.824160 489.664770 26.274692
0 arm_1 explore 3 2024-01-01 06:30:00 7.971751 8.612500 9.712263 25.837501 0.956747

        Rounds              : 101
        Optimum Reward      : 2188.713387428091
        Cumulative Reward   : 723.2076348563972
        Regret              : 1465.5057525716938

../_images/RL_mab_57_2.png

\(\epsilon\) = 0.4#

[33]:
model = epsilon_greedy(n_iterations=100, epsilon=0.4, arms=arms)
decision_idx decision seq ts arm reward reward_cumsum
0 1 explore 1 2024-01-01 00:30:00 arm_3 24.506117 24.506117
1 2 exploit 2 2024-01-01 01:00:00 arm_3 11.573501 36.079618
2 3 explore 3 2024-01-01 01:30:00 arm_2 7.527144 43.606761
3 4 exploit 4 2024-01-01 02:00:00 arm_3 47.611542 91.218303
4 5 explore 5 2024-01-01 02:30:00 arm_1 4.784452 96.002755
... ... ... ... ... ... ... ...
96 52 exploit 97 2024-01-03 00:30:00 arm_3 4.865530 563.582487
97 52 exploit 98 2024-01-03 01:00:00 arm_3 -20.385035 543.197452
98 53 explore 99 2024-01-03 01:30:00 arm_2 11.271790 554.469242
99 54 exploit 100 2024-01-03 02:00:00 arm_3 36.958096 591.427338
100 54 exploit 101 2024-01-03 02:30:00 arm_3 -8.002698 583.424640

101 rows × 7 columns

[34]:
experiment_rewarddf = model.rewardsdf
get_metrics(experiment_rewarddf)
plot_reward(experiment_rewarddf);
arm decision ts_count ts_min reward_min reward_mean reward_max reward_sum reward_std
3 arm_3 explore 12 2024-01-01 00:30:00 -54.355821 18.335123 65.663022 220.021473 32.191539
2 arm_3 exploit 67 2024-01-01 01:00:00 -44.863765 2.845487 74.965554 190.647654 25.756657
1 arm_2 explore 12 2024-01-01 01:30:00 2.282932 7.773443 11.271790 93.281310 2.549957
0 arm_1 explore 10 2024-01-01 02:30:00 3.860798 7.947420 14.250761 79.474202 3.907964

        Rounds              : 101
        Optimum Reward      : 1851.8474016216946
        Cumulative Reward   : 583.4246400457904
        Regret              : 1268.4227615759041

../_images/RL_mab_60_2.png