fermilab-accelerator-ai / meetings

Meeting notes and information
2 stars 0 forks source link

Online training - networked options #23

Open gnperdue opened 4 years ago

gnperdue commented 4 years ago

Suppose we attempt to train online in tandem with some networked PC - how do we collect and communicate experience (to a replay buffer style system, for example)? How do return gradients? (Do we rebuild the FPGA and re-flash or write to registers)? What is the right timing for that?

Likely some overlap with this issue and #12

gnperdue commented 4 years ago

It will also be important to weigh this against #22

A key issue will be making the system robust enough to work for production operations.