New LUGL framework enables gradient-boosted trees for reinforcement learning games

Researchers have developed a new framework called LUGL (Local Updates, Global Learning) that allows non-incremental learners, such as gradient-boosted trees (GBTs), to be effective in reinforcement learning (RL) settings. Traditionally, neural networks have dominated RL due to their ability to handle the non-stationary nature of self-play training. However, LUGL decouples data collection from model fitting, enabling GBTs like LightGBM to be trained on game states, which are inherently tabular. This approach alternates between local updates for accumulating game data and a global learning phase to train a generalizable function approximator. Experiments across various perfect and imperfect information games show that LUGL-based agents achieve competitive or superior performance compared to established methods like DQN and DeepCFR, challenging the prevailing reliance on neural networks for game-playing AI. AI

IMPACT Challenges the dominance of neural networks in game-playing AI and suggests gradient-boosted trees can achieve competitive or superior performance.

RANK_REASON The item describes a new framework and experimental results for applying gradient-boosted trees in reinforcement learning, presented in a research paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• Connect 4x4
• DeepCFR
• Flop5 Hold'em
• Goofspiel
• Hugging Face
• Kuhn's poker
• Leduc Hold'em
• Liar's Dice
• LightGBM
• Neural Networks
• Othello
• Tic Tac Toe

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →