Notebook Tutorials and Demonstrations

This is a collection of tutorial notebooks and demonstrations for the self-driving-lab-demo! First, you should get started with the main, public-facing tutorial. Just click the following “Open in Colab” badge:

Remotely Access a Free Self-driving Lab Demo

Traditional DoE vs. Bayesian Optimization

Multi-objective Optimization

Continous Multi-fidelity Optimization

Discrete Multi-fidelity Optimization

Multi-task Optimization

Batch Optimization

High-dimensional, Multi-Objective Optimization with Existing Data

Predefined Candidates

Benchmark Dataset Generation

All Tutorials

Note

Tutorials from sections 1, 2, and 3 are deprecated due to changes in hardware and software design of the demo (i.e., dropping “Blinkt!” in favor of built-in RGB LED on Maker Pi Pico base, and dropping a web server interface in favor of MQTT). Since they are instructive and show parts of the behind-the-scenes development process, they are kept here for reference and provenance.

1. Blinkt! Getting Started (deprecated)

• 1.0-sgb-blinkt-as7341-basic.ipynb

  • Let’s flash the LED and print out the sensor data!

2. Search Algorithms using Blinkt! (deprecated)

• 2.0-random-search.ipynb

  • 🚗 Let’s run a test drive of 100 random search iterations! 🚗

• 2.1-bayesian-optimization-blooper.ipynb

  • 💥Bayesian optimization is worse than random search and grid search… Wait what?💥

• 2.2-sensor-simulator.ipynb

  • 🕵️ Time to troubleshoot! Running simulations can help us to troubleshoot the source of the discrepancy. SPOILER: Oh! It was an issue with data processing 🤦 (but was that all? 🤨)

• 2.3-bayesian-optimization.ipynb

  • 🔁 Back to the algorithm comparison experiments! Lo and behold, Bayesian optimization is the most efficient. 😌

3. Pico W with a Web Server (deprecated)

• 3.1-random-vs-grid-vs-bayesian.ipynb

  🥑 Algorithm comparison using the Pico W that’s running a local web server 🥑

• 3.2-random-vs-grid-vs-bayesian-simulator.ipynb

  🥑 Algorithm comparison using a vamped up simulation based on the NeoPixel (as opposed to DotStar) LED 🥑

IoT-style control of Pico W

• 4.0-paho-mqtt-colab-sdl-demo.ipynb

  📡 Control the Pico W remotely using MQTT (The Standard for IoT Messaging) 📡

• 4.1-paho-mqtt-colab-sdl-demo-search.ipynb

  🔁 Run the same algorithm comparison experiments controlled from the cloud! 🔁

• 4.2-paho-mqtt-colab-sdl-demo-test.ipynb

  🌎 Remotely access a free, public test demo from anywhere in the world 🌍

Wired control of Pico W

• 5.0-nonwireless-search.ipynb

  No PicoW? No problem! 🤖 Run the same algorithm comparison experiments using a nonwireless Pico! 🤖

Advanced optimization

• 6.0-olympus-benchmarking-basic.ipynb

  🏋️ Let’s run some benchmarking experiments using the Olympus simulator! 🏋️

• 6.1-multi-objective.ipynb

  🎯 When more than one property is important (meaning virtually all real-world tasks 😉) 🎯

• 6.2-multi-fidelity.ipynb

  📈 Optimization using multiple information sources of varying cost and quality 📈

  • 6.2.1-multi-fidelity-continuous.ipynb

    📈 Optimization using continuous fidelities (e.g., runtime of a molecular dynamics simulation) 📈

  • 6.2.2-multi-fidelity-discrete.ipynb (WIP)

    📈 Optimization using discrete fidelities (e.g., finite number of allowed grid sizes in a simulation) 📈

  • 6.2.3-multi-task.ipynb (WIP)

    📈 Optimization using multiple, correlated information sources (e.g., simulations and experiments) 📈

• 6.3-batch-optimization.ipynb

  📦 Optimization using batches of samples. Don’t forget to condition your batches! (handled automatically via Ax platform 😁) 📦

Data ecosystem

• 7.0-data-ecosystem.ipynb (WIP)

  📊 Overview of notebooks in this section 📊

• 7.1-mongodb-read.ipynb

  📊 Read data from MongoDB 📊

• 7.2-robust-data-logging.ipynb (WIP)

  📊 Robust data logging 📊

• 7.2.1-hivemq-openssl-certificate.ipynb

  📊 Generate an OpenSSL certificate for HiveMQ 📊

• 7.3-benchmark-dataset-generation.ipynb

  📊 Generate a benchmark dataset 📊

eScience 2022

See also a set of tutorials prepared for the eScience 2022 conference. Video tutorials corresponding to these notebooks are published on YouTube in Taylor Sparks’ Optimization playlist.

• 1.0-traditional-doe-vs-bayesian.ipynb

• 2.1-gpei_hartmann_loop.ipynb

• 2.2-tune_cnn.ipynb

• 2.3-gpei_hartmann_service.ipynb

• 2.4-raytune_pytorch_cnn.ipynb

• 2.5-multiobjective_optimization.ipynb

• 2.6-continuous-multi-fidelity.ipynb

• 2.7-discrete_multi_fidelity_bo.ipynb

• 2.7.1-multi_task.ipynb

• 2.8-ax_service_existing_data.ipynb

  • see also 2.8.0.* notebook variations (e.g., SAASBO, MOO, batch, etc.)

• 2.8.1-human_in_the_loop.ipynb

• 2.8.2-ax_service_existing_data_saasbo_multi_objective.ipynb

• 2.9-scheduler.ipynb

• 2.10-paho-mqtt-colab-sdl-demo-test.ipynb

• 2.11-predefined-candidates.ipynb

More to come!

• discrete multi-fidelity optimization (simulation and experiments)

• high-dimensional Bayesian optimization (SAASBO, MORBO)

• scalable Bayesian optimization (MORBO, Dragonfly)

• asynchronous/batch optimization using network of experiments

• Grid search vs. random vs. Sobol vs. stochastic gradient descent vs. genetic algorithm vs. Bayesian optimization (e.g. via Olympus benchmarking platform)

• Repeat experiments for high-variance or failure-prone experiments via RayTune Repeater

• Combinations of above

• External evaluation of simulation functions (e.g. FuncX, Slurm, AWS, Google Cloud)

• Experimental orchestration software via Bluesky

• Storing experiments in a database backend (e.g. SQL, MongoDB)

• Combinations of above

Any requests? Post on the issue tracker 😉