Bayesian Optimization Accelerator
summary
An appliance-based toolkit running on an AC922 appliance (RHEL) exposes a minimal HTTP SDK and uses Gaussian processes with many kernel and acquisition choices, batch sampling and tuner settings to optimize expensive noisy black-box functions. Example connectors run shell or HPC jobs (STREAM benchmark) and a 13-parameter SPEC AMMP design space of 4,000 points. BOA finds the global optimum typically by ~25 iterations (median over 200 runs) and shows up to 138x fewer iterations versus grid search, with small variance across trials.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:35What is IBM Bayesian Optimization Accelerator
- 6:57Acquisition Functions
- 12:32Final Result after 15 Iterations
- 13:17BOA Architecture
- 18:41BOA Job Script Skeleton
- 26:01Objective Function - Analytic
- 26:51Objective Function - Look-up Table
- 28:35Objective Function - External Process
- 31:49BOA Job Configuration Parameters
- 34:53Three Types of Input Parameters
- 35:26The Use Case
- 37:11Parameter Space
- 37:57The Raw Data
- 38:42BOA Result Median Case
- 41:10Convergence Comparison
related talks
28mBayesian Optimization- Fundamentals, Implementation, and Practice
24mUsing Bayesian Optimization to Tune Machine Learning Models
48mBayesian Optimization of Gaussian Processes with Applications to Performance Tuning
39mBayesian optimisation with scikit-learn
01h 30mBayBE: A Bayesian Back End for Experimental Planning in the Low-To-No-Data Regime
41mOptimizing Machine Learning Models
34mBayesian Optimization and it's application to Neural Networks
26mAddressing Some of Challenges When Optimizing Long-to-Train-Models