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})

K = 10

def bandit(a: int, bias: np.array):
	zn = np.random.normal() + bias[a]
	return zn


def test(_episode_len: int, _qu: np.array, _nu: np.array, _bandit: Callable, _param: tuple[float, float]) -> np.array:
	_epsilon = _param[0]		# Anti-greediness (ability to explore)
	_rho = _param[1]		# reduce epsilon with age
	_r_mean = []

	# 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, 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
		_r_mean.append(r)

	return np.array(_r_mean)


if __name__ == '__main__':

	# Init parameters
	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(-1.5, +1.5, K)

			# Init Q and N
			qu = np.zeros(K)
			nu = np.zeros(K)
			r_mean += test(episode_len, qu, nu, _bandit=lambda a: bandit(a, ql_star), _param=param)

		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. bandits: K={K}")
	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, 1.5])

	pl.show()