ReCoML@MLSys '20

Is quantization helpful in the case of resource-constrained RL?

Here we show the benefits of quantizing a policy could lead upto 18x speedups! The table below shows results for various policy architectures trained for an end-to-end navigation task in Air-Learning that are trained finally and deployed at various precision levels on a Ras-pi 3b+. The figure below the table shows memory usage for Policy III

Checkout the Poster for more details!