qhjqhj00 beam-control-eval
FastML Science Benchmarks: Accelerating Real-Time Scientific Edge Machine Learning — Duarte et al. (2022) (arXiv:2207.07958, 2022)
What this evaluates
Evaluates real-time reinforcement learning for particle beam steering on edge SoCs. It probes the ability to maximize control reward while adhering to strict millisecond latency and cycle-rate constraints.
Datasets
- Fermilab Booster Synchrotron Dataset — total ?; splits: test (-1); repo https://github.com/fastmachinelearning/fastml-science
Metrics
Reward (R)(primary) — range: other- R = -|ΔI_min|, defined as the negative of the minimum error with respect to the reference expected current in the Booster.
Input / output format
Input: 5 expert-selected causal variables (32-bit floating-point) representing synchrotron and downstream accelerator currents/errors at each time step.
Output: 7 discrete actions output by a Deep Q-Network (DQN) agent, representing control signals for magnet power supplies.
Scoring recipe
# Per episode or cycle
delta_I_min = abs(current - reference_current)
reward = -delta_I_min
# Maximize cumulative reward over the episode
# Report reward convergence alongside latency (≤5ms) and pipeline interval (5ms)
Common pitfalls
- The environment is a surrogate LSTM model, not the physical accelerator; results may not transfer directly.
- Latency must include data movement overhead; the algorithm itself must run within 5 ms.
- Inputs are already pre-selected expert variables, not raw sensor data.
Evidence (verbatim from paper)
The primary performance metric in this reference benchmark is the reward, R, defined as the negative of the error with respect to the reference expected current in the Booster, R=−|ΔI_min|.
Citation
@misc{duarte2022fastml,
title={FastML Science Benchmarks: Accelerating Real-Time Scientific Edge Machine Learning},
author={Duarte et al. (2022)},
year={2022},
note={arXiv:2207.07958}
}
- arXiv: 2207.07958