Meta-Reinforcement Learning

To get a first grap of how the two-step task environment works, here is a gif of a meta-RL agent interacting with the environment for 100 trials (what we’ll call later an episode). For example, if after consistently choosing (which leads to ending up in with probability 80%) for 10 trials in row, the agent received 10 positive rewards in a row ( at the end of each trial), then is very likely to be the often-rewarded state. A useful way to think about A2C is as part of a broader class of algorithms in RL, called Actor Critic algorithms, where a neural network outputs both an actor and a critic:

Let’s suppose that, last trial, the agent chose the action , ended up in , and received a reward of ( was the often rewarded state).

Source: blog.floydhub.com