The point of TRPO is to try to find the largest step size possible that can improve the policy, and it does this by adding a constraint on the KL divergence ...
An actor-critic algorithm is a policy gradient algorithm that uses function estimation in place of empirical returns $G_t$ in the policy gradient update.
In order to use continuous action spaces and have stochastic policies, we have to model the policy $\pi$ directly. We can parametrize our policy using some p...
The original formulation of $Q$-learning requires that both the state and action space be small and discrete. However, we run into problems when the action s...
In the general reinforcement learning paradigm, the agent only has access to the state, the corresponding reward, and the ability to choose an action.
Every action of a rational agent can be thought of as seeking to maximize some cumulative scalar reward signal.
Neural style transfer is a technique for doing style transfer, where we combine the content of one image with the style of another image.
Autoencoders are a special kind of neural network used to perform dimensionality reduction. We can think of autoencoders as being composed of two networks, a...
Something funny happens when you take computer science. You begin to think of problems like this in terms of algorithms and mathematically optimal strategies.
