Richard Goodwin, Sesh Murthy, John Rachlin, and Rama Akkiraju
Effective production scheduling is at the heart of any efficient manufacturing process and is the key to profitability. Because scheduling is complex, manufactures use computers to generate near optimal schedules. However, using a computer to find a single near optimal schedule requires a precise definition of optimality which must tradeoff the competing interests of production efficiemy, customer satisfaction, profitability and product quality. Performing these tradeoffs automatically is beyond the fidelity of any utility model that could be constructed with reasonable effort. Scheduling systems work best as assistants to human schedulers, keeping track of complex constraints and presenting the a set of alternatives that highlight the tradeoffs. The system must allow the human scheduler to examine and manipulate the schedules in order for the human to arrive at the best schedule. The utility model must capture the high level factors that contribute to a good schedule, but it does not need to capture all the subtlety needed to select the best schedule. For manufacturing interests, scheduling objectives fall along four broad dimensions: Time, factors that relate to on time delivery, Quality, factors related to product quality, Money, factors related to profitability and Disruptions, factors related to the ease of manufacture. These dimensions map to the interests of the customer service representatives, quality engineers, accountants and manufacturing supervisors respectively. The human scheduler is presented with a set of schedules that represent the Pareto-Optimal frontier and negotiates a selection between these competing points of view. In this paper, we elaborate on our approach to interactive decision-support and illustrate it with examples from the IBM Load Planning System, which uses an asynchronous team of agents to create schedules for shipping manufactured goods to customers.