Unverified50% confidenceFactExact time
The main result of the paper is that Q-Learning with zero initialization and constant step size can, with high probability, reduce the sup-norm error to within ε using Õ(H^4/(μ_min·ε^2)+h_star) state transitions (for 0<ε≤1).
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe ε-greedy policy selects actions randomly with probability ε and selects the action with the highest current Q-value with probability 1-ε; ε is typically set to 1.0 at the beginning of training and then linearly decayed to 0.01 or 0.1 over millions of frames.70% similarUnverifiedState perturbation uses Gaussian noise injection (sigma=0.005) to prevent the model from learning fragile 1-dimensional linear trajectories.70% similarUnverifiedo1在某个任务中最终分数低于Claude 3.5，体现了强推理模型的'过度推理'问题——初始阶段出现偏差后会沿错误路径深入70% similarUnverifiedThe residual structure makes F=0 the default option—as long as the F(x) weights are close to zero, the residual block degenerates into an identity mapping, thereby resolving the optimization difficulty of SGD in learning identity mappings.69% similarUnverifiedε-prediction is prone to numerical instability at large embedding dimensions, because when α_bar_t approaches zero the scaling factor 1/sqrt(α_bar_t) tends to infinity, amplifying prediction errors.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1002187API
curl https://kongchang.com/api/v1/knowledge/claims/1002187MCP
get_claim(id=1002187)