How to formulate discounted return in cartpole?

Question

I am trying to formulate a problem that aims to prolong the lifetime of the simulation, the same as the Cartpole problem. I aware that there are two types of return:

• finite horizon undiscounted return (used for episodic problems)

$G = \sum_{t=0}^T R_t$

• infinite horizon discounted return (used for non-episodic problems).

$G = \sum_{t=0}^\infty \gamma^t R_t$

However, I'm confusing that "Is Cartpole episodic task?". Ideally, the simulation lasts forever. This is my final objective (prolonging the lifetime). But it still has some termination states. Should I introduce the termination state and use it with a discounted return like:

$G = \sum_{t=0}^T \gamma^t R_t$

Answer 1

It's not true that finite horizon MDP can't use a discount factor. Any discount factor $\gamma \leq 1$ is fine for finite horizon.

Whether or not cartpole is a finite or infinite horizon depends on the environment implementation. The default environment in openai gym uses $T=200$ as the horizon. This is customizable, and most papers I see which use cartpole as a test environment use $T=1000$ instead. The task is considered "solved" when the agent can reach $T$ timesteps consistently (e.g., the last 100 episodes have all reached $T$ timesteps).

If you want to treat cartpole as an infinite horizon task, that's fine too. Clearly you need $\gamma < 1$ here, and you can just $T$ to some arbitrary large number.