(428d) A Multi-Agent and Distributed Cloud Computing Approach for Industrial Production Scheduling Model Development and Deployment

Production scheduling [1] in Dow is a complex process which involves designing production cycles for multiple production trains. A production cycle is a repeating sequence of products that satisfies forecasted demand and balances inventory and product transition costs. Different product cycle sequences and campaigns result in different transition costs. Operationally, these cycles are designed using an excel-based tool and are unrolled into complete schedules which are then adjusted dynamically week-to-week in response to actual customer demand. This process relies heavily on experienced schedulers who build up tribal knowledge over time but also experiences frequent role turnover making the retention of this knowledge a challenge for the business. Furthermore, the current technology does not handle demand allocation of products that are shared across several production trains - this is left to the expertise and judgement of the subject matter expert schedulers. The current work process produces numerous inefficiencies in terms of excess inventory levels and transition costs.

In this talk, we will present a multi-agent [2] based mathematical optimization approach that allocates forecasted demand across multiple production trains and finds the detailed product cycle schedule. The overall optimization model is divided in two independent agents. The role of agent 1 is to determine optimal product demand allocations across the trains and ensure that the allocated products can be cycled through (i.e., the product graph is strongly connected [3]) to ensure feasible product transitions in the production cycle. The agent 1 iteratively loops through solving a MILP based product allocation model and a depth first search (DFS) algorithm [4] to check if the product solution is strongly connected. If the product allocation solution is not strongly connected, its corresponding binary solution is cut off in the next iteration until a strongly connected product solution is found. The role of agent 2 is to use the resulting product demand allocations from agent 1 and determine the product cycle design (such as product sequence and campaign length) and production schedule for each production trains. Simultaneous product cycle design and scheduling is computationally expensive. Agent 2 thus leverages a hybrid MILP scheduling and simulation-based approach to address the high computational complexity. Agent 2 first runs a sampling algorithm to generate candidate product sequences and campaign lengths, which are then solved in the scheduling model to determine the inventory and transition loss cost. The product cycle schedule with the least cost is then implemented in the production site.

Both the demand allocation model (agent 1) and the single train detailed schedule model (agent 2) are deployed to Dow’s Azure [5] platform. These are deployed as independent Azure Machine Learning [6] pipeline endpoints. Once these endpoints are called a batch compute cluster is spun up, the models are solved, and the results returned. The demand allocation model will trigger the single train model when the optimal demand allocations are calculated without returning to the user between agent 1 and agent 2 solves. A user may also choose to execute only the single train model when they determine that a multi-train model is unnecessary. Agent 2’s sampling solution approach is distributed across the compute cluster using Anyscale’s Ray [7] package to solve each sample in parallel thus greatly reducing the solution time to that of a single sample solve time which is on the order of seconds.

The models/agents were tested during a pilot phase deployment where scenarios from different production trains are run and evaluated, and model updates are made to capture all complexities. Having successfully completed this pilot, efforts are underway to scale the solution to over 50 production trains world-wide by the end of the 2022.

This work demonstrates the power of combining distributed and parallel cloud computing technologies with traditional MILP models with a sampling approach which may have been impracticable for sequential solution approaches. Furthermore, the deployment and management of the independent agents is done with the help of MLOps [8] strategies which allows for the seamless and automatic testing and deployment of model updates to end users. The approach shows great promise and potential for expansion to include additional agents, potentially using additional methods like machine learning and artificial intelligence, into this multi-agent system where agents trigger other agents automatically and as needed.

References:

[1] Harjunkoski, Iiro, Christos T. Maravelias, Peter Bongers, Pedro M. Castro, Sebastian Engell, Ignacio E. Grossmann, John Hooker, Carlos Méndez, Guido Sand, and John Wassick. "Scope for industrial applications of production scheduling models and solution methods." Computers & Chemical Engineering 62 (2014): 161-193.

[2] Wooldridge, Michael. An introduction to multiagent systems. John wiley & sons, 2009.

[3] Nuutila, Esko, and Eljas Soisalon-Soininen. "On finding the strongly connected components in a directed graph." Information processing letters 49, no. 1 (1994): 9-14.

[4] Awerbuch, Baruch. "A new distributed depth-first-search algorithm." Information Processing Letters 20, no. 3 (1985): 147-150.

[5] Collier, Michael, and Robin Shahan. Microsoft Azure Essentials-Fundamentals of Azure. Microsoft Press, 2015.

[6] Klein, Scott. "Azure machine learning." In IoT Solutions in Microsoft's Azure IoT Suite, pp. 227-252. Apress, Berkeley, CA, 2017.

[7] https://www.anyscale.com/ray-open-source

[8] Demystifying XOps: DataOps, MLOps, ModelOps, AIOps and Platform Ops for AI. Gartner Research. 2021