3 problems

0 votes0 replies0 views
Linear convergence conjecture for RMSProp and Adam on the softmax policy gradient objective
Let the softmax policy gradient objective be the objective considered in the paper, and let $\texttt{RMSProp}$ and $\texttt{Adam}$ denote the corresponding optimization methods. RM…
Conjectureopenadaptive-methods optimization policy-gradientsarXiv posted 2 days ago
0 votes0 replies0 views
Conjecture on the noise sensitivity of policy-gradient estimates
Noise-sensitivity conjecture. The gradient estimates are more sensitive to noise than the discounted rewards.
Conjectureopenpolicy-gradients reinforcement-learning stochastic-optimizationarXiv posted 2 days ago
0 votes0 replies0 views
Conjecture on the bounded monotonicity of policy-gradient oracle variance
Variance monotonicity conjecture. Under proper assumptions, the variance of the policy-gradient oracle is an increasing but bounded function of $\tau$ .
Conjectureopenpolicy-gradients reinforcement-learning stochastic-optimizationarXiv posted 2 days ago

Linear convergence conjecture for RMSProp and Adam on the softmax policy gradient objective