Does any one know any example code of an algorithm Ronald J. Williams proposed in
A class of gradient-estimating algorithms for reinforcement learning in neural networks
Yes, do a search on GitHub, and you will get a whole bunch of results:
The most popular ones use this code (in Python):
__author__ = 'Thomas Rueckstiess, [email protected]'
from pybrain.rl.learners.directsearch.policygradient import PolicyGradientLearner
from scipy import mean, ravel, array
class Reinforce(PolicyGradientLearner):
""" Reinforce is a gradient estimator technique by Williams (see
"Simple Statistical Gradient-Following Algorithms for
Connectionist Reinforcement Learning"). It uses optimal
baselines and calculates the gradient with the log likelihoods
of the taken actions. """
def calculateGradient(self):
# normalize rewards
# self.ds.data['reward'] /= max(ravel(abs(self.ds.data['reward'])))
# initialize variables
returns = self.dataset.getSumOverSequences('reward')
seqidx = ravel(self.dataset['sequence_index'])
# sum of sequences up to n-1
loglhs = [sum(self.loglh['loglh'][seqidx[n]:seqidx[n + 1], :]) for n in range(self.dataset.getNumSequences() - 1)]
# append sum of last sequence as well
loglhs.append(sum(self.loglh['loglh'][seqidx[-1]:, :]))
loglhs = array(loglhs)
baselines = mean(loglhs ** 2 * returns, 0) / mean(loglhs ** 2, 0)
# TODO: why gradient negative?
gradient = -mean(loglhs * (returns - baselines), 0)
return gradient