Any Tips for per-player tracking? about alphazero.jl HOT 6 CLOSED

What you propose is a good solution and the way I would do it by default.

I would rather them not be aware of the other players score

What do you mean by this exactly? If you want the value prediction from the network to be made independently from the other player score, you can just remove this information in the GI.vectorize_state function.

Another question: are future available actions dependent on both players' score? If not, you can remove both scores from the state and just use rewards to model score. If this info is relevant, there is no choice but to include it in the state.

from alphazero.jl.

Hm ok. Let me think on that.

Another question: are future available actions dependent on both players' score? If not, you can remove both scores from the state and just use rewards to model score. If this info is relevant, there is no choice but to include it in the state.

As for this kinda?
Each can make one of two moves lets say: Jump, wait, land
So each should only be able to make the moves Jump and wait, but then if they've jumped they should only be able to wait or land.
I also need to retain what the value of the map was when they jumped in order to calculate their score when they landed.

So how best would you handle the actions and the reward portion of it? Would you perhaps go through the previous moves and use that to determine if they had already jumped as well as use that to add up the values needed to determine which performed better?

from alphazero.jl.

By definition of a state, which future scenarios are possible should only depend on the current state of your game and never on the history of previous actions. (Technically, this is called the Markovian property of states.)

Therefore, in your case, the state should contain some info about whether or not the agent is currently "in the air" and waiting to land.

from alphazero.jl.

Hm ok.
Can you elaborate on a few things for me?
For Vectorize state
-Can you explain how this is used by alphazero?
-What would you do if you needed to vectorize a large amount of data?

from alphazero.jl.

vectorize_state turns a state into a format that can be fed into a neural network (typically a tensor of floating point numbers).

What would you do if you needed to vectorize a large amount of data?

I do not understand your question. If your states are large and yield big tensors, then it is what it is. Either you can find more economical representations or you have to live with your system being memory-hungry and maybe stressing your hardware.

from alphazero.jl.

I am closing this issue for lack of activity but do not hesitate to reopen it if you need more help.

from alphazero.jl.