IBM BOA for POWER
- 1. What is Bayesian
Optimization?
Bayesian optimization is a sequential design strategy for
global optimization.
Many workflows require you to find a powerful set of
parameters solve a problem. The challenge is finding those
parameters robustly in as little time as possible.
© 2019 IBM Corporation
Applied to Computational Chemistry
Applied to Engineering Design
Applied to Drug Discovery
BOA accelerated workflow uses
1/3 of the calculations to
achieve 4 orders of magnitude
resolution increase
BOA performed in 19 hours and ~30 simulations what an
expert designer would do in 3 weeks
Brute force methods of
screening require 20,000
experiments. BOA accelerated
method required ~200
IBM BOA gets HPC clients designs here
faster than any other method
- 2. Blackbox function optimization
§ No derivatives f’(x).
§ No analytic form.
§ Possibly multiple minima.
§ Possibly noisy data.
§ Expensive to calculate.
§Grid search - exhaustive
§Random search - luck
§Simulated annealing – large number of
evaluations
§Gradient descent - requires derivatives
§Genetic algorithm – hard to tune
Bayesian Optimization
- 3. Air flow simulation around a car,
FPGA synthesizer, reservoir
simulation, etc.
GPU accelerated Power server (IBM AC922)
High performance computing infrastructure
(x86/Power Systems/Cloud etc.)
Parameter values
Result (Chip power
consumption, for example)
Bayesian optimization
1. Pick a surrogate that
represents your prior belief
on black box behavior.
2. Define an acquisition
function over the surrogate.
3. Repeat:
• Select the parameter values
by optimizing the acquisition
function.
• Evaluate the black box for
those parameter values.
• Update the surrogate
through posterior inferencing
- 4. © 2019 International Business Machines Corporation CONFIDENTIAL – INTERNAL USE ONLY
Interface Functions
Input Data
Output Data
SOLVE
Traditional HPC
BOA
Interface
Function
(out)
Interface
Function
(in)
NEW: Interface Functions
Scheduler
Objective
Function
User Defined
Unique to each Client
- 5. IBOA
IBM BOA Differentiation
• Dimensionality mitigation based on compressive
sensing
• Smart initialization
• Novel acquisition functions
• Explainability
• Ease of use through software abstractions
- 7. Explore-Exploit Trade-Off
• Exploration – Prefers to acquire new knowledge
• Exploitation – Prefers to lean heavily on what is already known to drive
optimization
• Need to strike a balance – too much exploration is inefficient, too much
exploitation can result in poor performance.
Exploitation PI, 𝜀 = 0
Exploration PI, 𝜀 = 0.2
Probability of Improvement
Probability of Improvement does not consider the amount by which an improvement occurs, just
that an improvement occurs.
Can we tweak this to tell the algorithm what constitutes a significant improvement
𝛾 =
𝑓 𝑥!"#$ − 𝜇 𝑥 + 𝜀
𝜎(𝑥)
Denote the improvement as
𝛾 =
𝑓 𝑥!"#$ − 𝜇(𝑥)
𝜎(𝑥)
Probability of Improvement (PI) (Kushner):
𝛼%& = Φ(𝛾 𝑥 )