Also called Derivative Free Optimisation ![[Pasted image 20230521193836.png]] When there is a function that we cannot access but we can only observe its outputs based on some given inputs, it is called a **black-box function.** On the other hand, black-box optimization (BBO) deals with optimizing these functions. Characteristics: - Structure of the objective function is not known - Constraints defining the set is unknown, unexploitable or non-existent - Does not have an algebraic model that can be solved analytically - You may evaluate it at given points and from these results you may derive and use any information you want Eg. optimization problem where you need to conduct simulations to evaluate your solution. ![[Pasted image 20230521193846.png]] ![[Pasted image 20230521193901.png]] ### Application Domains - UI/UX - A/B testing - Machine Learning Hyper-parameterisation - Physical Design - Simulation based Shape Optimisation - Robotics - Chemistry - Design of photofunctional molecules and medical drugs - optimisation of thermal emission materials - high Li-ion conductive solid electrolytes - discovery of a new phase in inorganic thin films for solar cells - Material Science ![[Pasted image 20230521193917.png]] More here: [https://www.sciencedirect.com/science/article/pii/S2192440621001386](https://www.sciencedirect.com/science/article/pii/S2192440621001386) [Automated discovery, design, and optimization based on black-box optimization](https://pubs.acs.org/doi/10.1021/acs.accounts.0c00713#) ### Practical Considerations - Minimize Configurations possible for end-users - Names, Parameters, Access Control Lists (ACLs) - Integration w/ Jupyter Notebook / [Colaboratory](https://colab.research.google.com/) - Scalability - ![[Pasted image 20230521193928.png]] - Reliability - Testing, Monitoring, Enabling Introspective and Defensive measures - Modularity of the Suggestion Workers - Having state of the art algorithms - GP Bandits - SMAC - CMA-ES - Simulated Annealing - Genetic Algorithms - Random Search - Grid Search - ![[Pasted image 20230521193941.png]] ### Current Solutions We thus use numerical methods of optimization or population based approaches that need only the value of the fitness function. - Bayesian Optimisation - Reinforcement Learning - Active Learning ### Market Insights - Vizier - Black Box OptimisationTool from Google - [https://research.google/pubs/pub46180/](https://research.google/pubs/pub46180/) - Open Source version: [https://github.com/google/vizier](https://github.com/google/vizier) - [https://arxiv.org/pdf/2106.00421.pdf](https://arxiv.org/pdf/2106.00421.pdf) - [https://www.ini.rub.de/PEOPLE/glasmtbl/projects/bbcomp/](https://www.ini.rub.de/PEOPLE/glasmtbl/projects/bbcomp/) ![[Pasted image 20230521193957.png]] ### Challenges TBC ### Benchmarking Functions that are cheap to evaluate, but have various kinds of properties. Ref - No Free Lunch Theorem. See: [https://numbbo.github.io/workshops/](https://numbbo.github.io/workshops/) ![[Pasted image 20230521194015.png]] ![[Pasted image 20230521194032.png]] Dimensions impact performance ![[Pasted image 20230521194050.png]] Ref: [[Bayesian Optimisation]] #### Psuedo code of Teaching Learning based Optimisation ![[Pasted image 20230521194109.png]] --- References: 1. [https://www.youtube.com/watch?v=6aSG8SdvkoU&ab_channel=GoogleDevelopers](https://www.youtube.com/watch?v=6aSG8SdvkoU&ab_channel=GoogleDevelopers) 2. [https://www.lix.polytechnique.fr/~dambrosio/blackbox_material/Cassioli_1.pdf](https://www.lix.polytechnique.fr/~dambrosio/blackbox_material/Cassioli_1.pdf) 3. [https://towardsdatascience.com/machine-learning-how-black-is-this-black-box-f11e4031fdf](https://towardsdatascience.com/machine-learning-how-black-is-this-black-box-f11e4031fdf)