@inproceedings {pub4627,
	title = {An Empirical Analysis ofMeasure-Valued Derivatives for Policy Gradients},
	author = {Fabio Muratore},
	year = {2021},
	abstract = {AUTHORS: Jo ̃ao Carvalho, Davide Tateo, Fabio Muratore, and Jan Peters

The success of policy gradient methods in the context of Reinforcement Learning, particularly in robotic tasks, lies in computing a precise (low variance) and accurate (low bias) gradient estimate.
Traditional policy gradient algorithms obtain a gradient estimate by employing the likelihood-ratio trick, which is known to produce unbiased but high variance estimates.
In this work, we study how a different type of stochastic gradient estimator, the Measure-Valued Derivative, can be applied to Reinforcement Learning tasks.
This estimator is unbiased, has low variance, and applies to differentiable as well as non-differentiable functions. We empirically evaluate this estimator in step-based policy gradients and show that it can reach comparable performance with methods based on the likelihood-ratio or reparametrization tricks, both in low and high-dimensional action spaces.},
	publisher = {IEEE},
	booktitle = {International Joint Conference on Neural Networks (IJCNN)}
}