Seems to be almost identical to the previous paper for ICLR 2022: Learning by Directional Gradient Descent (by Silver et al.) (https://openreview.net/pdf?id=5i7

Seems to be almost identical to the previous paper for ICLR 2022: Learning by Directional Gradient Descent (by Silver et al.) (https://openreview.net/pdf?id=5i7lJLuhTm https://openreview.net/pdf?id=5i7lJLuhTm) The paper has some additional experimental analysis of the properties of the algorithms (including how many samples you should estimate per epoch), and also extends the method for training RNNs or reinforcement learning.