How can reinforcement learning be applied when the goal location or environment is unknown?

Question

I am studying RL. I was thinking whether a new state value or the observation is provided by the environment before the agent actually implements the action.

Take the maze problem as an example. Each state consists of all the available cells information, provided by the environment. But what if the environment is unknown? For example, there is a maze with an unknown destination cell. The agent needs to find the destination cell. The state is 1 or 0, meaning the destination reached or not. But the environment, which is the maze, can only provide the state at cell $i$ which is 0 or 1 only when the agent reaches cell $i$.

Can this still be solved by RL? I am confused about the environment setup.

Answer 1

When reading about RL and RL agents/algorithms, you always need to keep in mind that, typically, the RL agent/algorithm is trying to maximize the reward (or something equivalent, such as minimizing the regret) in the long run (i.e also the reward that you may receive in the future): that's its (mathematical) goal. Whether that also corresponds to the high-level goal (e.g. reaching some physical location in some world) that the (human) designer of the so-called reward function (i.e. the function that gives the reward to RL agent) had in mind is a different story.

To maximize the reward, the RL agent interacts with the environment by taking actions, receiving rewards, and moving to other states. Initially, the RL agent does not know which actions lead to more rewards, so it may take random actions (this is known as exploration). Once it starts to understand the dynamics of the environment, it may start to take only the actions that lead to high reward (this is known as exploitation).

To answer your question directly, the RL agent can indeed take actions without knowing the dynamics of the environment. However, initially, it may need to take some random actions (which may lead to low rewards), so that to get more rewards in the long run.