Today, IBM is releasing a new appliance: IBM Bayesian Optimization Accelerator, designed to help product and design teams introduce products and features faster by lowering their design time to generate new and better products.
The problem spaces that product innovators are working with are getting more complex. We find our clients are demanding faster answers, forcing teams to re-examine current best practices and consider alternative methods because “building bigger” won’t be an option forever. Clients tell us that internal stakeholders are demanding that solutions come faster, cost less and be more accurate than ever. Meanwhile, budget isn’t expanding to meet all the new needs.
In a perfect world, what would solve all of these challenges?
-
Methods that do not require knowledge about a problem beforehand.
-
The ability to make the most of infrastructure by parallelizing efforts and spending less CPU and time getting to the ultimate answer.
-
Easy-to-use services which can cope with high dimensionality because the name of the game is optimizing real problems, not simple academic ones.
-
Traceable and non-biased methodology, particularly in regulated industries.
Simply put, the ideal solution for these rising challenges would deliver fast innovation and superior results while using fewer resources.
Introducing a more efficient way: IBM Bayesian Optimization Accelerator
With IBM Bayesian Optimization Accelerator, a state-of-the art general parameter optimization tool created based on cutting-edge innovations from the IBM Research team, users only need to define design variables, objective and constraints to leverage a powerful optimization engine. This is an appliance and can be accessed as a full solution from IBM–including hardware, software and installation services.
Simply put, IBM Bayesian Optimization Accelerator is designed to find the optimal solution for complex, real-world design problems in less time and using less resources.
-
Fast innovation: Bayesian Optimization Accelerator is designed to find solutions quickly with easy, quick initial integration and methods that require fewer starting inputs and scale in parallel to help decrease the time to result.
-
Superior results: Bayesian Optimization Accelerator locates the most optimum solution 89% of the time with traceable, explainable optimization decisions that require no prior data, avoiding bias.
-
Fewer resources: These methods can be applied without specialized data science skills and make existing infrastructure more efficient, keeping costs down while still responding to business needs.
The solution works by helping the HPC cluster know where to look. It sits outside of the traditional HPC cluster and is dedicated to running Bayesian Optimization methods only. The HPC cluster will send the values for constraints and objective functions to the appliance, which will send back new locations in the search space to find optimal solutions.
What makes this solution different?
Bayesian methods are not new to the mathematical world. However, we have found that standard and freely available Bayesian methods, such as Greedy or Monte Carlo search, suffer from several challenges that make applying a Bayesian search methodology to a product design problem challenging and often impractical.
IBM Bayesian Optimization Accelerator can scale to orders of magnitude number of dimensions which allows it to tackle real world problems instead of simplistic ones. And unlike search methods such as IBM implementations of Greedy and Monte Carlo searches, it determines design points with much fewer samples required, which allows it to get to results faster and cheaper. In fact, in comparison tests run by IBM Research, Bayesian Optimization Accelerator reached the least regret solution in the fastest time in over 82% of IBM’s tested experiments against IBM implementations of Greedy and Monte Carlo algorithms.
In those same comparison tests, this solution delivered the answer with the least regret over 89% of the time. To do this, it does not require any prior knowledge of a design problem. With a proprietary “bootstrapping” method, it can start an optimization from no initial data, gather initialization data on its own, and then start the Bayesian optimization process.
Furthermore, Bayesian Optimization Accelerator provides a graphical optimizer “explainability” interface for users who are interested in traceability of model design history and optimizer choices to help build trust in the methodology. This means that scientists can interrogate the optimizer during an experiment about why it chose to evaluate suggested parameters.
IBM is already pursuing application areas for this technology across many industries, such as aerospace, automotive, electronic design and oil and gas. In electronic design, IBM’s own signal integrity research team used Bayesian Optimization Accelerator to reduce the time required to reduce their signal integrity simulations by 99.3%, from nearly eight days to just 80 minutes.Meanwhile in oil and gas, a research team sliced their time to reach results by 61% by using Bayesian Optimization Accelerator to identify the ideal mix of injection materials and timing into reservoirs to maximize their outputs.
These and other early results with teams across the globe indicate that IBM Research’s work in the lab is already driving real results for businesses and helping them deliver superior results faster and using fewer resources.
FIND OUT MORE ABOUT IBM BAYESIAN OPTIMIZATION ACCELERATOR & BOOK TO TEST AT BOSTONLABS
[1] Based on IBM internal testing performed by IBM Research on 771 sample sets of drug discovery data obtained from the ChEMBL database, which is licensed under Creative Commons Attribution-ShareAlike 3.0. (https://creativecommons.org/licenses/by-sa/3.0/legalcode). Tests run using IBM Bayesian Optimization Accelerator on one IBM Power Systems AC922 server. Optimization initialization performed with industry-standard chronological approach. Results valid as of 25 September 2019 and conducted under laboratory conditions. In this testing, IBM Bayesian Optimization Accelerator correctly identified the least regret solution 89% of the time (687/771 data sets). [2] Based on IBM internal testing performed by IBM Research on 771 sample sets of drug discovery data obtained from the ChEMBL database, which is licensed under Creative Commons Attribution-ShareAlike 3.0. (https://creativecommons.org/licenses/by-sa/3.0/legalcode). Tests compare IBM Bayesian Optimization Accelerator run on one IBM Power Systems AC922 server to comparable Greedy and Monte Carlo algorithms implemented by IBM, based on industry best practice and run on one Power System S822LC for High-Performance Computing. Optimization initialization performed with industry-standard chronological approach. Results valid as of 25 September 2019 and conducted under laboratory conditions. In this testing, IBM Bayesian Optimization Accelerator reached the least regret solution in the fastest time when compared to aforementioned algorithms 82% of the time (344/420 experiments conducted). [3] Based on IBM internal testing during POWER10 development comparing IBM Bayesian Optimization Accelerator run on one IBM Power System AC922 server to traditional brute force ‘design of experiments’ methodology implemented by IBM, based on industry best practice and run on non-accelerated x86 Linux architecture. Results valid as of July 2018. In testing, traditional methods required 11260 minutes to reach final result, and IBM Bayesian Optimization Accelerator required 80 minutes (99.3% reduction). [4] Based on currently pre-published research conducted by Dr. Mary Wheeler and Xueying Lu of the University of Texas at Austin Oden Institute for Computational Engineering and Sciences. IPARS simulations run using IBM Bayesian Optimization Accelerator on one Power System S822LC for High-Performance Computing and compared to industry-standard genetic algorithms authored by researchers and run on the same server. Results valid as of October 2020 and conducted under laboratory conditions. Research simulated a reservoir using compositional Cranfield model and optimized how to best inject CO2 and surfactant into the reservoir. Using IBM Bayesian Optimization Accelerator, the team reached equivalent result as genetic algorithm methods (objective function value of 1.31e8 for BOA vs. 1.29e8 for genetic algorithm) in 61.5% fewer iterations (73 vs. 200 iterations).