Files
k-bandits/k-bandits.py
T
2025-04-02 15:55:52 +02:00

84 lines
2.1 KiB
Python

import numpy as np
import matplotlib.pyplot as pl
from typing import Callable
float_formatter = "{:.3f}".format
np.set_printoptions(formatter={'float_kind': float_formatter})
def bandit(a: int, bias: np.array):
zn = np.random.normal() + bias[a]
return zn
def test(_k_arms: int, _episode_len: int, _bandit: Callable, _param: tuple[float, float]) -> np.array:
_epsilon = _param[0] # Anti-greediness (ability to explore)
_rho = _param[1] # reduce epsilon with age
_r_list = []
_a_list = []
# Init Q and N
_qu = np.zeros(_k_arms)
_nu = np.zeros(_k_arms)
# Calc z in advance
z = np.random.uniform(size=_episode_len)
# Calc a_expl in advance
a_expl = np.random.randint(low=0, high=_k_arms, size=_episode_len)
for j in range(0, _episode_len):
_a = a_expl[j]
# choose action
if z[j] > _epsilon:
# Choose best action
_a = np.argmax(_qu)
# get reward from bandit
_r = _bandit(_a)
_nu[_a] = _nu[_a] + 1
_qu[_a] = _qu[_a] + (_r - _qu[_a]) / _nu[_a]
# Reduce tendency to explore with number of steps (or with age for humans)
_epsilon = _epsilon * (1 - _rho)
# Statistics
_a_list.append(_a)
_r_list.append(_r)
return np.array(_r_list), np.array(_a_list)
if __name__ == '__main__':
# Init parameters
k_arms = 10
highest_reward = 1.5
num_realisations = 2000
episode_len = 1000
params = [(0.0, 0.0), (0.01, 0.002), (0.1, 0.002)]
legend = []
for param in params:
# Init r_mean
r_mean = np.zeros(episode_len)
for realization in range(0, num_realisations):
# init bandits with different biases for shifting reward probability
# -> expected reward q*(a)
ql_star = np.random.uniform(-highest_reward, +highest_reward, k_arms)
r_list, a_list = test(_k_arms=k_arms, _episode_len=episode_len, _bandit=lambda a: bandit(a, ql_star), _param=param)
r_mean += r_list
pl.plot(r_mean/num_realisations)
legend.append(f"Param {param}")
pl.grid()
pl.title(f"Norm. E(R), num. realizations: Z={num_realisations}, num. bandit arms: K={k_arms}")
pl.xlabel("Episode")
pl.ylabel("E(R)/Z")
pl.legend(legend)
ax = pl.gca()
ax.set_xlim([-episode_len/10, episode_len])
ax.set_ylim([0, highest_reward])
pl.show()