Original title: Local Updates, Global Learning (LUGL): Playing Games with non-incremental Learners
David Milec, Spyridon Samothrakis, Michael Fairbank, Dennis J. N. J. Soemers · September 2026
Neural networksComputer systems inspired by the brain that learn patterns from data. dominate reinforcement learningA way for computers to learn by trying actions and seeing rewards, like training a pet., a method where computer programs learn by playing games against themselves. They do this because they can learn step by step as new data arrives. However, decision tree algorithms like LightGBM are often better, faster, and state of the art for tabular data, meaning data arranged in neat rows and categories like cards or board squares. Trees usually struggle in reinforcement learning because the incoming data constantly shifts. To fix this, researchers created a framework called LUGL, short for Local Updates, Global Learning. It separates collecting game data from training the model. First, during a local phase, an agent plays games against itself and stores game values in a finite table. Next, during a global phase, the model trains on that table to learn patterns for unseen states before the table resets. Across nine tested board and card games, tree-based agents matched or beat standard neural network methods.
Imagine you're trying to master a complex board game by playing it repeatedly. If the game rules or your strategy keep changing as you play, you might never settle on a good approach because each new game feels different. In this paper, the computer faces similar shifts while learning to play games, but tree-based models can't adjust during play. LUGL solves this by having the computer play and record key moments in a notebook, then study that notebook to improve, so it can handle changes without updating on the fly.
This matters because it challenges the default use of neural networks in game AI, showing that simpler, faster tree-based methods can work just as well. This could save computational resources and improve performance in real-world applications like game development and AI research.
The abstract doesn't state specific limitations, but it focuses on testing in specific games and compares only to DQN and DeepCFR. It doesn't address how LUGL might perform in more complex, real-time environments or with other types of non-incremental learners.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Local Updates, Global Learning (LUGL): Playing Games with non-incremental Learners — David Milec, Spyridon Samothrakis, Michael Fairbank et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.