My vague understanding of reinforcement learning (RL) is that it's very similar to supervised learning except that it updates on a continuous feed of data/activity, this to me sounds very similar to AutoML (which I've started to notice being used).
Do they use different algorithms? What is the fundamental difference between RL and AutoML?
I'm after an explanation for somebody who understands technology but does not work with machine learning tools regularly.
Automated machine learning (AutoML) is an umbrella term that encompasses a collection of techniques (such as hyper-parameter optimization or automated feature engineering) to automate the design and application of machine learning algorithms and models.
Reinforcement learning (RL) is a sub-field of machine learning concerned with the task of making decisions and taking actions in an environment so as to maximize (long-term) reward (which is the goal of the so-called RL agent). RL is (at least partially) based on the way animals (including humans) learn. For example, the usual way of training a dog to perform a certain task is to reward it with food whenever it takes the correct action (for example, jumping, if you want the dog to jump whenever you make a certain gesture with your hand). In this case, the RL agent is the dog, the task the dog needs to perform (e.g. jumping) is the environment, food is the reward and the goal is to get food.
Given that reinforcement learning (RL) is a sub-field of machine learning, then, in principle, AutoML can also be used to automate the design of RL algorithms, models or agents. For example, if you use a neural network to represent the policy (the function the determines which action to take in the environment), then you can potentially use AutoML to find the most appropriate architecture (for example, the most appropriate number of layers) for this neural network.
RL can be used in the context of Neural Architecture Search (NAS), with is a form of automated ML. A model searches for an architecture that performs a given task. How well this task is performed guides how the architecture will be modified (improved) on the next pass. It works but is very computation-intensive (think hundreds of GPUs).
See for instance:
