An intelligent agent-based framework to design warehouse management systems that assist human operators in dynamic demand environments to mitigate the impact of exceptions: a simulation

The increase in e-commerce and omni-channel commerce is having a significant impact on the supply chain sector and its warehouses. Although automation is becoming prevalent as the solution to increasing warehouse complexity, many warehouses cannot automate due to product unsuitability and/or prohibitive capital expenditure. The golden rule of material handling is smooth product flow, but there are exceptions that occur in the day-to-day work of a warehouse that can have significant effects on product flow and impact customer satisfaction. The supply chain literature on risk and event management deals with the impact of an exception or “disruptive event” on the supply chain and how it can ripple along the supply chain if not contained at its source. A major research gap is that exceptions within the warehouse and their potential for disruption are not understood. Warehouse operations involve a sequence of material handling processes, each with various associated costs. Strategies to reduce the cost of these processes focus on optimising them along the metrics of time, quality, and productivity. The warehouse process optimisation literature includes storage assignment algorithms, pick path strategies, optimal warehouse layouts and task allocation. However, optimisation strategies mostly operate on the perfect scenario and do not account for the impact of exceptions on the successful completion of a process and the associated cost of the rework. This research focuses on how warehouse management systems can be designed to be more robust to exceptions that can cause a disruption, augmenting the human decision maker with intelligent decision support. The purpose was to design a distributed, real-time, agent-based model for warehouse management that could detect, mitigate, and resolve exceptions while aiding the human operator with a digital mentor. This research uses Design Science Methodology. The problem identification and motivation phase of the project was qualitative. This included:1) An exploration of common warehouse issues via exploratory interviews with warehouse practitioners. Through thematic analysis, there emerged ten main warehouse themes: automation, decision support, e-commerce, labour (people), management, manual operations, process, supply chain, systems, and warehouse exceptions. The warehouse themes provided the context within which the exception categories were grounded. The exception categories were data, operational issues, resource issues, unexpected events and the human intervention required to resolve exceptions. Further analysis of the transcripts relating to exceptions resulted in a ripple effect caused by exceptions, and three categories of resulting disruption: 1) Disruption to warehouse efficiency, 2) Disruption to warehouse productivity and 3) Disruptions to the supply chain. When the exceptions were condensed further, the following high-level underlying causes emerged: Operator attention lapses, operator cognitive limits, small problems being ignored, time-consuming and complicated resolution processes, lack of visibility, lack of critical information, lazy scheduling of priorities, some tasks are faster but less accurate when done off-system, process workarounds to fix system gaps, and a cost vs accuracy trade-off mentality. An agent-based model for the design of warehouse management systems called HAN/DSL/MENT (HDM) was proposed. The HDM model is composed of real-time intelligent and communicating agents in a HAN (Human Agent Network). The three types of agents are entity agents, service agents and mentor agents (MENT). Entity agents are digital twins of non-human warehouse entities. Service agents provide services or perform a management role within the network. Mentor agents are differentiated from the other digital agents because they are always paired with a human operator and interact in the HAN on behalf of the human operator. Mentor agents augment the ability of the human operator to perform their job by assisting with task completion and decision support.  The DSL (Decision Support Layer) which is composed of service agents, is responsible for high-level continuous trend analysis and providing appropriate algorithms and process execution priorities to meet the environmental conditions and constraints of the environment. A simulation of the model was developed using JADEX and tested with 10 common warehouse exception condition scenarios. The evaluation of the HDM model involved simulation runs in two modes: traditional and mentor assisted. A comparison between modes was completed using parametric and non-parametric statistical analysis tests. The ten test scenarios were exception scenarios and dealt with the exception sub-categories of congestion, picking exceptions, replenishment exceptions, abandoned pallets, inventory management, and maverick operators.  Mentor-assisted mode was better than traditional mode at a statistically significant level for measures within the dimensions of time, quality, and productivity. The results of this research show that real-time intelligent software agents that work in tandem with a human operator within a human-agent system can improve the quality/accuracy of warehouse operations by improving warehouse performance measures within the categories of time, quality, and productivity. This is achieved by improvements to the visibility of warehouse exceptions, targeted troubleshooting and resolution assistance and adaptive scheduling of priorities. This allows time-critical information to be used in real-time enabling time-critical decisions and actions to be taken in proactive and reactive modes for short-term decision support. In addition, a decision support layer enables incremental changes to be made via long-term decision support driven by trends analysis. A network of real-time intelligent agents working to mitigate the occurrence of warehouse exceptions and effectively and quickly resolve them can also have a positive impact on labour productivity and motivation. Labour productivity and motivation can be enhanced by spending less time on rework (after a task fails), with implicit checks (reducing cognitive load) and with guided and targeted exception trouble shooting and resolution support. The HDM model proposes a different approach to designing warehouse management systems. This non-centralised model consists of a human-agent network within which three components of digital agents communicate. The first component is autonomous agents connected to non-human entities and service agents. The second component is digital mentor agents assisting human operators and decision-makers. The third component is a decision support layer involved in more complex data analysis and decision support. The HDM model is not specific to warehouses but can be applied to other complex, dynamic environment where humans and systems (computers) must interact to achieve a common goal. Warehouses can be considered one of the cases of its application.