Improving Robustness of AlphaZero Algorithms to Test-Time Environment Changes
This addresses robustness issues for AlphaZero users in dynamic or uncertain environments, but it is incremental as it builds on existing methods.
The paper tackles the problem of AlphaZero agents failing when test environments differ from training, and shows that simple modifications to the standard framework significantly boost performance, even with low planning budgets.
The AlphaZero framework provides a standard way of combining Monte Carlo planning with prior knowledge provided by a previously trained policy-value neural network. AlphaZero usually assumes that the environment on which the neural network was trained will not change at test time, which constrains its applicability. In this paper, we analyze the problem of deploying AlphaZero agents in potentially changed test environments and demonstrate how the combination of simple modifications to the standard framework can significantly boost performance, even in settings with a low planning budget available. The code is publicly available on GitHub.