When n is independent ofδ our approach yields an algorithm whose sample complexityconvergesto 2n2 log 1δ asngrows.

This work proposes a middle ground in the form of a scalable information-theoretic generalization of CCA, termed max-sliced mutual information (mSMI).

It is unclear whether distributed errors alone are sufficient for synapses to make coordinated updates to learn complex, nonlinear reward-based learning tasks.

TRPO, PPO), and a critic that involves minimizing a closely connected objective.

Empirically, we verify the computational efficiency of our methods.

This phenomenon cannot be easily explained by offline RL's return maximization objective.