AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

RAST-MoE-RL: A Regime-Aware Spatio-Temporal MoE Framework for Deep Reinforcement Learning in Ride-Hailing

Yuhan Tang, Kangxin Cui, Jung Ho Park, Yibo Zhao, Xuan Jiang, Haoze He, Jiangbo Yu, Haris Koutsopoulos, Jinhua Zhao

ICML 2026 · Reinforcement learning › Visual and representation RL · approach: RL for combinatorial routing · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace the 12M plain MLP encoder with the 12M MoE (16,4) encoder under PPO (ablation)PPO policy · SF Uber trajectories (held-out test)matching wait (s): 208 -> 182Improved
replace the 12M plain MLP encoder with the 12M MoE (16,4) encoder under PPO (ablation)PPO policy · SF Uber trajectories (held-out test)pickup wait (s): 577 -> 523Improved
replace a 48M dense Transformer with a 12M MoE (16,4) encoder (ablation)PPO policy · SF Uber trajectories (held-out test)pickup wait (s): 545 -> 523Improved
replace a 12M dense Transformer with a 12M MoE (16,4) encoder (ablation)PPO policy · SF Uber trajectories (held-out test)matching wait (s): 201 -> 182Improved
use top-4 routing instead of top-2 with 16 experts (ablation)MoE PPO · SF Uber trajectories (held-out test)pickup wait (s): 543 -> 523Improved
scale to 32 experts (top-8, 24M) from 16 experts (top-4, 12M) (ablation)MoE PPO · SF Uber trajectories (held-out test)matching wait (s): 182 -> 188Hurt
scale to 64 experts (top-8, 48M) from 16 experts (top-4, 12M) (ablation)MoE PPO · SF Uber trajectories (held-out test)pickup wait (s): 523 -> 536Hurt
use 8 experts top-4 (6M) instead of 16 experts top-4 (ablation)MoE PPO · SF Uber trajectories (held-out test)pickup wait (s): 523 -> 541Hurt
use GRPO-style within-state normalization instead of PPO (ablation)MoE (16,4) · SF Uber trajectories (held-out test)pickup wait (s): 523 -> 528No clear effect
use PPO instead of A2C (ablation)MoE (16,4) · SF Uber trajectories (held-out test)pickup wait (s): 555 -> 523Improved
use PPO instead of DQN (ablation)MoE (16,4) · SF Uber trajectories (held-out test)pickup wait (s): 612 -> 523Improved
use adaptive reward instead of fixed ratio 8:1 (ablation)PPO 16 experts top-4 · SF Uber trajectories (reward ratio, PPO 16 experts top-4)pickup wait (s): 1746 -> 524Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy