Getting started with Bonsai: Phase 2 – Delivery & real-world trials
Neal Analytics takes an end-to-end lifecycle approach to Project Bonsai engagements. It’s a three-phase process that goes from pre-engagement to delivery and real-world trials, and finally to scaling.
But first, let’s have a quick recap on what Bonsai is, and what it’s used for…
What is Bonsai?
Microsoft Project Bonsai is an AI toolchain that can be used to control and optimize production systems. It leverages Deep Reinforcement Learning (DRL) and Machine Teaching to create an AI agent, or a “brain”, that can provide human operators or control systems with the best decisions available to achieve an optimization goal.
We call these Autonomous Systems, which are well suited to complex industrial systems and manufacturing processes. Learn more about Project Bonsai and Autonomous Systems here.
Moving from pre-engagement to delivery
In my last blog, I wrote about Phase 1 of a Bonsai project, also known as the pre-engagement phase. Here, our team works with business stakeholders and SMEs to identify and prioritize value-creating use cases, clarify challenges, and map current processes. We then design an AI agent (aka Bonsai “brain”) that we propose and present to the client.
Building a simulated environment
Simulation modelling is a digital representation of a physical process, where one can make changes to key parts of the process to see what could happen. This capability enables businesses to run “What If scenarios”, one of the early proof points within a typical Bonsai engagement.
A Bonsai brain will leverage a simulation model to accelerate the process of running hundreds of million scenarios to learn about specific environments states and determine the best decision(s). This will allow the brain to optimally control the overall targeted process.
In the early stages of the delivery phase, the team will focus on either leveraging/validating existing client simulations or defining and building a simulation model. Third party simulation software vendors will generally be leveraged to accelerate the build stage. A simulation build can take up to 3 months to develop.
Brain training and real-world trials
The team will then gradually transition to brain training and trials. Here is where we leverage Deep Reinforcement Learning (DRL) models to train a brain; the brain learns through practice (trial and error) in the simulated environment, and good decisions will be rewarded accordingly.
Machine Teaching enhances the models by injecting subject matter expertise into the training process – this helps accelerate the learning process for the brain and brings transparency to an “AI black box”.
Once the brain has finished learning in the simulated environment, we prepare the solution to be deployed on a real test line where operators will validate the Bonsai brain’s decisions and execute accordingly. This process will continue to further refine the brain before it goes to production.
Once the brain is in production, it only needs to look around, check the state of the live environment, and make a decision. Then, after a given time interval, it looks around again at the state of the environment, and makes the next right decision, and so on, and so on.
Completing Phase 2
Delivery and real-world trials can take between 6 to 9 months to complete. Upon completion of this phase, the Neal team will start planning the final phase – rolling out the brain to multiple lines / locations that have similar characteristics.
In my next blog, I’ll cover the key elements of Phase 3: Scaling Bonsai solutions.
Want to learn more about Project Bonsai and the best practices for training a brain? Contact us.
- Getting started with Project Bonsai: Phase 1 – Pre-Engagement
- Want to learn more about DRL, Machine Teaching, and Autonomous Systems? Check out our explainer video series.
- Microsoft – Learn more about Project Bonsai
- Step-by-step guide: Getting started with Project Bonsai
- Customer story: How PepsiCo leveraged Autonomous Systems to create more perfect Cheetos