When meeting with customers today there is inevitably a conversation about the applications and use cases that are best suited for the Bonsai Platform. And while coming at it from different angles, both parties are well aligned with the intent of the discussion: 1) Can we identify a specific hardware or software application that would drive increased business value from greater intelligence, 2) Is the Bonsai Platform well suited to enable you to program AI models to deliver on this objective?
With Bonsai initially optimized for a particular set of applications, specifically programming control and optimization into industrial systems including robotics, supply chain, manufacturing and HVAC, it is extremely important to be very targeted when identifying appropriate use cases for the platform. Our primary objective in any customer engagement is to help identify the ‘best fit’ use cases as quickly as possible. To help expedite this discovery process, we have devised three questions to qualify whether or not a specific application is a fit for Bonsai:
1. Is there an application/system that would see increased business value from greater automation and/or operational efficiency? Before going any further in a customer engagement it is important to confirm that the application in question falls within Bonsai’s current area of focus on the AI use case spectrum.
While there are many possible use cases that would benefit from increased intelligence, Bonsai is optimized today for programming AI models that improve system control and enhance decision support. Building greater intelligence into these applications results in lower costs, and increased operational efficiency.
2. Can the use case be solved with deep reinforcement learning? This one is a bit tricky. The Bonsai Platform currently uses deep reinforcement learning algorithms to train AI models. Deciphering whether or not an application can increase automation, or enhance operational efficiency through deep reinforcement learning requires a better understanding of a how the system interacts with its environment, and how that interaction changes the state of the environment. The more dynamic the interactions are between the controller (e.g. a robot) and its environment, the better suited reinforcement learning is relative to data-centric techniques.
To better understand how exactly a system interacts with its environment we usually walk through this sequence of questions below:
a. Does the system interact with a real world environment? In this question we are trying to gauge if a system's predictions, or actions in the case of a robot, impact state, or have material consequences to the environment it interacts with? If the answer is yes, this suggests that the system may be a good candidate for RL.
b. Is the environment the system interacts with always presented consistently, or does it vary? In the case of a specific part, item, or package, is it in the same place and orientation all the time? If the environment varies frequently this implies that the system may be a good candidate for RL.
c. Are the system's interactions with its environment predictable or unpredictable? For example, if a self-driving car is instructed to turn left at an intersection, and the same arrangement of cars is present at that intersection as has been previously observed, you cannot expect the behavior of the environment to be the same as before. If the interactions are unpredictable, this means that the system may be a good candidate for RL.
d. Does system setup need to be re-tooled often to accommodate new runs/configurations? A real life example is the factory restructuring that takes place to accommodate a new vehicle model. If retooling occurs frequently then the system may be a good candidate for RL.
3. Are simulations a viable training source for this project? With the Bonsai Platform today, learning is greatly enhanced from interacting with a simulation of a real world environment. As a result, it is important to understand whether or not a system's operations and interactions with its environment can be simulated, or modeled. An example of this would be an HVAC system, which can be both modeled in advance, while also collecting real-time telemetry.
Working through these questions helps our customers quickly determine whether or not there is an application that could be programmed with greater intelligence using our AI Platform. Is there a hardware or software system within your business that aligns the criteria outlined above? If so, please contact us to talk through your use case in greater detail, or apply directly for our early access program here.