I dream of agent playing 2048 using this library~

[LinXueyuanStdio]

May it come true?

[puyuan1996]

Hello,

Thank you very much for your attention. Yes, we are confident that this can be achieved. The Stochastic MuZero algorithm is designed to solve environments with stochasticity in dynamics, such as 2048. We are currently developing the 2048+Stochastic MuZero algorithm, and have completed about 80% of the work. We will submit the related PR in a few weeks, so please stay tuned. Thank you for your patience.

Best wishes.

[PaParaZz1]

We are working on the implementation of Stochastic MuZero with 2048 env in #64. You can keep track of it.

[LinXueyuanStdio]

Yeah, I have been following the PR for days.

[PaParaZz1]

#77 shows some experiment results on 2048 environment.

[puyuan1996]

• Hello, we have comprehensively optimized the 2048 environment and the related configurations for Bot, MuZero, and Stochatic MuZero in this PR. We are now providing the benchmark results in the 2048 environment as shown in the following image:

• Furthermore, we've provide the render video to illustrate the performance of Random action, Bot action, MuZero, and Stochastic MuZero agents in the 2048 environment.

  • Random action (~1000):
    

  • Bot action (~30000-70000):
    

  • MuZero sim=100 (~27956):
    

  • Stochastic MuZero sim=100 (~37356）:
    

  • Where the term sim represents the number of simulations used for evaluation in MCTS, and the score within brackets corresponds to the score procured by each respective agent.

Thank you for your attention.