Deep RL is hot these days. It's one of the most popular topics in the submissions at NeurIPS / ICLR / ICML and other ML conferences. And while the definition of RL is pretty general, in this note I'd argue that the famous REINFORCE algorithm alone is not enough to label your method as a Reinforcement Learning one.
This is the final post of the stochastic computation graphs series. Last time we discussed models with discrete relaxations of stochastic nodes, which allowed us to employ the power of reparametrization.
These methods, however, posses one flaw: they consider different models, thus introducing inherent bias – your test time discrete model will be doing something different from what your training time model did. Therefore in this post we'll get back to the REINFORCE aka Score Function estimator, and see if we can fix its problems.
Last year I covered some modern Variational Inference theory. These methods are often used in conjunction with Deep Neural Networks to form deep generative models (VAE, for example) or to enrich deterministic models with stochastic control, which leads to better exploration. Or you might be interested in amortized inference.
All these cases turn your computation graph into a stochastic one – previously deterministic nodes now become random. And it's not obvious how to do backpropagation through these nodes. In this series I'd like to outline possible approaches. This time we're going to see why general approach works poorly, and see what we can do in a continuous case.