Learning-Rate Clocks

Cross-source consensus on Learning-Rate Clocks from 1 sources and 5 claims.

1 sources · 5 claims

Risks & contraindications

In average-reward reinforcement learning, changing from local clocks to global clocks can affect stability. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
Average-reward TD analyses need to treat clock choice as part of the algorithm. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
Classical convergence proofs often index step size by visits to each state. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
Practitioners more commonly use step sizes indexed by elapsed time. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
In discounted reinforcement learning, the local-clock and global-clock distinction has not produced known divergence counterexamples under standard assumptions. — On the Divergence of Differential Temporal Difference Learning without Local Clocks