lightrl 0.0.1

A lightweight reinforcement learning package

lightrl

Lightweight Reinforcement Learning python library.

Multi-armed_bandit

Example

Typical example is when you want to call API, but you are being blocked. With this package you can automatically find the optimal number of requests that should be sent together in order to achieve error rate below certain treshold.

import time
import random
from typing import Tuple

from lightrl import EpsilonDecreasingBandit, two_state_time_dependent_process


class SimulatedAPI:
    def __init__(self):
        # Initialize variables to keep track of requests
        self.time_window_requests = []
        self.window_length = 1  # 60 seconds window
        self.request_limit = 200  # request limit in a window
        self.block_duration = 1  # block duration in seconds
        self.blocked_until = 0

    def request(self) -> Tuple[int, int]:
        current_time = time.time()

        # Remove requests older than the current window
        while (
            self.time_window_requests
            and self.time_window_requests[0] < current_time - self.window_length
        ):
            self.time_window_requests.pop(0)

        # Analyze if blocked
        if current_time < self.blocked_until:
            return 500

        if len(self.time_window_requests) > self.request_limit:  # Over the limit
            self.blocked_until = current_time + self.block_duration
            return 500

        self.time_window_requests.append(current_time)
        return 200


if __name__ == "__main__":
    api = SimulatedAPI()
    request_nums = [10, 25, 50, 100, 150, 200, 250, 300, 350, 400, 450, 500]
    bandit = EpsilonDecreasingBandit(
        arms=request_nums, initial_epsilon=1.0, limit_epsilon=0.1, half_decay_steps=100
    )

    def api_request_fun(request_num):
        success_cnt = 0
        fail_cnt = 0
        for _ in range(request_num):
            http_status = api.request()
            if http_status == 200:
                success_cnt += 1
            else:
                fail_cnt += 1
            time.sleep(0.0001)
        return success_cnt, fail_cnt

    two_state_time_dependent_process(
        bandit=bandit,
        fun=api_request_fun,
        failure_threshold=0.1,
        default_wait_time=0.1,
        extra_wait_time=0.1,
        waiting_args=[10],  # Working with 10 requests in the waiting state
        max_steps=1000,
        verbose=True,
        reward_factor=1e-6,
    )

This script will run for 3 mins and then it will output at the end

Q-values per arm:
  num_tasks=10: avg_reward=0.00010, count=6
  num_tasks=25: avg_reward=0.00019, count=12
  num_tasks=50: avg_reward=0.00030, count=15
  num_tasks=100: avg_reward=0.00087, count=7
  num_tasks=150: avg_reward=0.00083, count=11
  num_tasks=200: avg_reward=0.00113, count=109
  num_tasks=250: avg_reward=0.00010, count=11
  num_tasks=300: avg_reward=0.00010, count=13
  num_tasks=350: avg_reward=0.00009, count=16
  num_tasks=400: avg_reward=0.00010, count=11
  num_tasks=450: avg_reward=0.00013, count=9
  num_tasks=500: avg_reward=0.00010, count=17

Multi-armed bandit correctly found out that the optimal number of tasks is num_tasks=200.