007: Licence to sKill
Putting intelligent, trained agents into the industrial field
Cue the James Bond music… no, not that one, that’s Mission Impossible… yep, that’s the one.
Imagine a highly skilled intelligent agent, trained by expert teachers over every possible scenario and utilizing the best possible technology mixed with human initiative and insights that can be deployed to save the world (or at least optimize your industrial operations).
This is the (slightly twisted for dramatic effect) premise behind Composabl, the brainchild of Kence Anderson and his team who have built a platform for the design and development of intelligent autonomous agents which utilize machine teaching and AI to sense and perceive industrial operations and respond with intelligent, human-like decisions in real-time to control and optimize processes.
I first met Kence several years ago whilst he was running Microsoft’s Project Bonsai team which had acquired / developed a unique approach to industrial optimization through deep reinforcement learning (DRL) whereby simulations are utilized to train AI “brains” which can then be deployed in the field. The key difference between DRL and standard supervised machine learning is that while supervised learning requires a model to be trained on a “correct” answer output from each input, DRL learns the correct behaviors based on incremental rewards from moving in the correct optimization trajectory. I worked with Kence and his team on a number of initiatives and was always impressed by his ability to explain quite complex optimization problems, and the technologies/approaches with which to address them.
Bonsai’s core idea was to empower and enable discipline and operational experts, who understood their industrial process intimately but didn’t necessarily have extensive data science skills, to codify their knowledge and train autonomous systems with AI. Once these systems have been trained they can be deployed either as decision support systems, complementing existing control algorithms, or embedded as specific closed-loop controllers.
After a very successful independent start-up, Microsoft acquired Bonsai in 2018 based on their unique approach to industrial optimization using machine teaching. The Bonsai system used a low code environment to set up the simulation and training environment needed to effectively teach an AI system to respond to a particular physical or process problem with defined optimization objectives using expert knowledge and heuristics. Based on the success of the Bonsai approach, Microsoft elevated this approach to encompass “The Industrial Metaverse” in 2022.
Several of us at Pontem worked on Bonsai projects and whilst we found DRL to be very powerful in addressing some complex industrial problems we also noted some limitations where we were unable to integrate other aspects of control theory and approaches… it was DRL or nothing. But the promise was there to deliver optimization of some of the more tricky optimization problems which relied on the ability to understand the operational environment, weigh up competing objectives and utilize expert knowledge and operational heuristics to determine the best course of action for optimization.
In integrating Bonsai into the broad scope of their Industrial Metaverse vision, Microsoft had set the scene for rapid expansion of DRL in industrial processes with the might of their enterprise IT reach and the talent of the legacy Bonsai team…
Then Microsoft axed Bonsai at the start of 2023.
The combination of an overambitious goal of integrating an AI approach to address the whole of the “industrial metaverse” and the emergence of other AI technologies (GenAI) may have ultimately led to the shelving of the project. Through discussions with the Bonsai team at the time, we also shared a view that while DRL was useful in certain circumstances there were also a number of other traditional approaches which were also needed as part of a holistic toolbox to address a wide range of optimization problems. This lack of flexibility for a reinforcement learning only approach meant that the number of use cases able to be addressed was somewhat limited. I think this is what primarily led to the axing of the program.
But like James Bond emerging through the smoke from the rubble of a ruined building… Composabl was born! (too dramatic?).
I met up with Kence at AVEVA World to discuss his new venture and was quickly convinced that his direction for Composabl addressed all of the areas of improvement we experienced with Bonsai and that the ethos of the company matched Pontem’s approach seamlessly, combining the power of data science and machine learning with domain expertise to solve complex operational problems. Within a month we had signed a partnership agreement and have recently completed an intensive partner enablement training session.
As part of the partnership discussions we sat down with Kence to talk about Composabl and ask (00)7 questions on how intelligent autonomous agents will be part of the modern industrial operating environment.
Question 001: How did Composabl come about and how is it different from Bonsai/Microsoft?
Microsoft Project Bonsai was a powerful platform for creating industrial agents. It had a well-loved user interface but was a closed SaaS platform focused mostly on one AI technique (Deep Reinforcement Learning). Composabl enables engineers to build industrial-strength agents by combining existing automation techniques (like expert rules, control theory, and optimization algorithms) with advanced AI and machine learning as skills. Composabl is an open platform that allows various stakeholders (like data scientists and controls engineers) to contribute their algorithms and models to the agent training process.
Question 002: What are autonomous agents and how can they be used in modern industrial scenarios?
Intelligent Autonomous Agents control and optimize industrial processes by sensing and responding in real time. They exhibit more human-like decision-making than traditional automation. These agents provide five aspects of human decision-making that traditional automation cannot.
Question 003: What are the key steps to consider when developing autonomous agents?
Machine Teaching is a methodology for designing and building intelligent agents. Machine Teaching suggests that to design an agent, we should first identify the skills that we want to teach the agent, then orchestrate the skills in the agent, then determine which algorithm or technique should be used to execute each skill.
Question 004: How does deep reinforcement learning fit in with traditional control methods (Rules, PID, APC, MPC etc)?
Learning how to control a system or process by practicing is the best way to generate a controller that is very adaptable to changing situations and scenarios. This is true for humans and for artificial intelligence. For example, practice allows humans to learn nuances that they cannot learn from following simple rules. For this reason, Deep Reinforcement Learning (DRL), an algorithm that learns by practicing, is a great addition to agents that need to be adaptable to changing conditions. DRL is also quite good at blending, non-linear control, and navigating fuzzy boundaries. But DRL has its drawbacks. For example, it has a very difficult time obeying constraints. Every control and optimization technique has strengths and weaknesses. This is why agents that combine decision-making techniques well tend to outperform single techniques and traditional automation.
Question 005: How important is domain knowledge and industry experience in a successful autonomous agent deployment?
Controlling and industrial system (like a machine) or process (like a manufacturing line or logistics supply chain) is like navigating a complex landscape. Your destination is the goal of the process (like maximizing throughput or efficiency). Domain knowledge and industry experience represents experience in navigating this landscape. This expertise is crucial to design and build performant, explainable, agents that work in real life.
Question 006: How do autonomous agents fit in the broader discussions of AI and the overall AI hype cycle?
Intelligent Autonomous Agents make decisions. They “do something” to take action in a real environment. Many technologies in the AI hype cycle (like computer vision or digital twins or simulation), perceive something. They help you determine what is happening in the environment. Other technologies in the hype-cycle support AI but aren’t actually artificial intelligence. IOT is an example of this. Generative AI is extremely exciting, and potentially useful inside of intelligent agents, but certainly not trustworthy to make decision in real industrial environments.
Question 007: What are the benefits in collaborating with strategic partners to deliver successful projects?
System Integrator partners are crucial to deliver successful projects. They bring deep understanding of and experience with specific verticals and the system and processes within them. Composabl provides tools for strategic partners to design and build intelligent agents that revolutionize their specialty verticals.
Partner Enablement Workshops: Pontem is ready for the field
Over 2 days in February, key technical and leadership personnel from Pontem undertook a workshop on the utilization of the Composabl platform for the design and building of intelligent autonomous agents. The course covered the fundamentals of modular design of these agents for decision support and closed loop control of complex industrial operations.
It was an intense couple of days where we explored the limits of the platform and built some agents through machine teaching and integration of various control approaches from the initial stages of agent configuration and design through to integration of simulations for multiple scenario training… we need a montage!
Kence, Andy and Octavio delivered a fantastic couple of days of workshops and we were able to establish a great working relationship which will deliver some great outcomes for our clients.
"We had a highly engaging workshop with the Pontem Analytics team. They were able to quickly learn the Machine Teaching methodology for architecting Industrial Strength AI Agents then apply that knowledge to several customer scenarios.
The team was able to dive straight into the SDK, code up several Agent designs including Multi Skill Agents with Strategy Pattern, integrate a simulation to Composabl, train them all and evaluate their results with some benchmarks. This was a significant achievement and impressive how quickly they developed an end-to-end Intelligent Agent!! We are very much looking forward to working together on future customer engagements"
As part of the workshop Pontem are now certified by Composabl as Intelligent Agent Designers and Builders. We are officially licenced to skill! Keep an eye out for upcoming case studies where we will be putting this into action and sending our unique intelligent agents out into the field.
