MiniGrid and/or Maze environment support for AlphaZero

[YuriyPryyma]

Thank you for open-sourcing such a fantastic library!

I am trying to train my own environment using the LightZero library.
My environment has the following characteristics:

• Discrete action space
• Single-player
• Deterministic dynamics (with a simulator available)
• Sparse rewards
• Tensor-shaped state space

I believe the environment most similar to mine would be something like MiniGrid or Maze. In LightZero, only MuZero-like algorithms support these envs. However, these algorithms involve learning the dynamics, which is not ideal for me since I already have a deterministic environment simulator. Not learning environment dynamics would be more sample-efficient, so I want to use AlphaZero approach.

I noticed LightZero paper states that support for MiniGrid or Maze envs with AlphaZero is under development.

My questions:

1. Is there an intermediate, unmerged implementation of MiniGrid or Maze environments with AlphaZero that I could use as an example?
2. How can I generally utilize LightZero's AlphaZero implementation for single-player games?

[YuriyPryyma]

I have noticed some recent MR #245 that adds "single_player_mode"

But as it looks like AlphaZero is still very much focused on board games as policy code operates with "board" variable