How would you train a reinforcement learning agent from raw pixels? For example, if you have 3 stacked images to sense motion, then how would you pass them to neural networks to output Q-learning values?
A Convolutional Neural Network (CNN) structure is a standard neural network architecture when working with 2D pixel input in reinforcement learning, and it is the technique used in the original DQN paper (see paragraphs 1 & 3 of Section 4.1 of https://arxiv.org/abs/1312.5602). CNNs typically take 3-dimensional input, where the first two dimensions are
width of your images, and the third is
rgb color. The technique in the paper was to convert each RGB frame (or image) to greyscale format (so it has only 1 color channel/dimension instead of 3) and instead use the
rgb_color dimension as a
frames dimension that is indexed by each stacked frame.
Currently, I am watching a YouTuber: Machine Learning with Phil, and he did it very differently. On the 13th minute, he defined a network that outputs a batch of values rather than Q-values for 6 states. In short, he outputs a matrix rather than a vector.
Later in the tutorial series, he most likely will discuss the training of the neural network. During training, you need to find the q-values of a batch of sets of stacked frames. Specifically, each element of the batch is a set of stacked frames. In other words, a set of stacked frames is treated as a single observation, so a batch of sets of stacked frames is a batch of observations.
To find these q-values, you will perform a forward pass of the batch of observations through the neural network. A forward pass of a single observation (set of stacked frames) through the neural network yields a vector of q-values (one for each action). Thus, a forward pass of a batch of observations (batch of stacked frames) will yield a matrix of q-values (one vector of q-values for each observation (or set of stacked frames)). This technique is used because many standard neural network libraries are designed to perform a forward pass on a batch of inputs through the neural network much faster than performing a forward pass on each input separately.