A demo of Bandit BAI on a real dataset to select mirrors

In this exemple we use a sequential halving agent to find the best server to download ubuntu from among a choice of 8 french servers.

The quirck of this application is that there is a possible timeout when pinging a server. We handle this by using the median instead of the mean in sequential halving’s objective.

The code is in three parts: definition of environment, definition of agent, and finally definition of the experiment.

[INFO] 07:09:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.496    3           5
[INFO] 07:09:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.489    0           10
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.489    0           18
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.488    0           26
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.307    7           36
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.311    4           46
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.307    7           56
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.31     7           66
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.307    7           76
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.306    7           86
[INFO] 07:10:        agent_name  worker  reward  action  max_global_step
                         SH        0     -0.307    7           96
[INFO] 07:10: ... trained!
[INFO] 07:10: Saved ExperimentManager(SH) using pickle.
[INFO] 07:10: The ExperimentManager was saved in : 'rlberry_data/temp/manager_data/SH_2024-07-19_07-09-53_d7c15a5d/manager_obj.pickle'
The optimal action (fastest server) is server number  7

import numpy as np

from rlberry.manager import ExperimentManager, read_writer_data
from rlberry.envs.interface import Model
from rlberry_research.agents.bandits import BanditWithSimplePolicy
import rlberry.spaces as spaces

import requests
import matplotlib.pyplot as plt


import rlberry

logger = rlberry.logger

# Environment definition


TIMEOUT = 2

mirrors_ubuntu = np.array(
    [
        "https://ubuntu.lafibre.info/ubuntu/",
        "https://mirror.ubuntu.ikoula.com/",
        "http://ubuntu.mirrors.ovh.net/ubuntu/",
        "http://miroir.univ-lorraine.fr/ubuntu/",
        "http://ubuntu.univ-nantes.fr/ubuntu/",
        "https://ftp.u-picardie.fr/mirror/ubuntu/ubuntu/",
        "http://ubuntu.univ-reims.fr/ubuntu/",
        "http://www-ftp.lip6.fr/pub/linux/distributions/Ubuntu/archive/",
    ]
)


def get_time(url):
    try:
        resp = requests.get(url, timeout=TIMEOUT)
        return resp.elapsed.total_seconds()
    except:
        return np.inf  # timeout


class MirrorBandit(Model):
    """
    Real environment for bandit problems.
    The reward is the response time for French servers meant to download ubuntu.

    On action i, gives a negative waiting time to reach url i in mirror_ubuntu.

    WARNING : if there is a timeout when querying the mirror, will result in
    a negative infinite reward.

    Parameters
    ----------
    url_ids : list of int or None,
        list of ids used to select a subset of the url list provided in the source.
        if None, all the urls are selected (i.e. 8 arms bandits).
    """

    name = "MirrorEnv"

    def __init__(self, url_ids=None, **kwargs):
        Model.__init__(self, **kwargs)
        if url_ids:
            self.url_list = mirrors_ubuntu[url_ids]
        else:
            self.url_list = mirrors_ubuntu

        self.n_arms = len(self.url_list)
        self.action_space = spaces.Discrete(self.n_arms)

    def step(self, action):
        """
        Sample the reward associated to the action.
        """
        # test that the action exists
        assert action < self.n_arms

        reward = -get_time(self.url_list[action])
        terminated = True
        truncated = False
        return 0, reward, terminated, truncated, {}

    def reset(self, seed=None):
        """
        Reset the environment to a default state.
        """
        return 0, {}


env_ctor = MirrorBandit
env_kwargs = {}

# BAI Agent definition


class SeqHalvAgent(BanditWithSimplePolicy):
    """
    Sequential Halving Agent
    """

    name = "SeqHalvAgent"

    def __init__(self, env, **kwargs):
        BanditWithSimplePolicy.__init__(
            self, env, writer_extra="action_and_reward", **kwargs
        )

    def fit(self, budget=None, **kwargs):
        horizon = budget
        rewards = []
        actions = []
        active_set = np.arange(self.n_arms)

        logk = int(np.ceil(np.log2(self.n_arms)))
        ep = 0

        for r in range(logk):
            tr = np.floor(horizon / (len(active_set) * logk))
            for _ in range(int(tr)):
                for k in active_set:
                    action = k
                    actions += [action]
                    observation, reward, terminated, truncated, info = self.env.step(
                        action
                    )
                    rewards += [reward]
                    ep += 1
            reward_est = [
                np.median(np.array(rewards)[actions == k]) for k in active_set
            ]
            # We estimate the reward using the median instead of the mean to
            # handle timeout.
            half_len = int(np.ceil(len(active_set) / 2))
            active_set = active_set[np.argsort(reward_est)[-half_len:]]

        self.optimal_action = active_set[0]
        self.writer.add_scalar("optimal_action", self.optimal_action, ep)

        return actions


# Experiment

xp_manager = ExperimentManager(
    SeqHalvAgent,
    (env_ctor, env_kwargs),
    fit_budget=100,  # we use only 100 iterations for faster example run in doc.
    n_fit=1,
    agent_name="SH",
)
xp_manager.fit()

rewards = read_writer_data([xp_manager], preprocess_tag="reward")["value"]
actions = read_writer_data([xp_manager], preprocess_tag="action")["value"]


plt.boxplot([-rewards[actions == a] for a in range(6)])
plt.xlabel("Server")
plt.ylabel("Waiting time (in s)")
plt.show()

print(
    "The optimal action (fastest server) is server number ",
    xp_manager.agent_handlers[0].optimal_action + 1,
)