Scenario A, there is a notable tail end where the MCTS solver performs slightly worse. The 10th percentile of greedy-guided MCTS performance falling only slightly below 100% shows that ggMCTS performs nearly on par with the greedy benchmark even in the worst of cases. Table 7 shows that the greedy benchmark often struggles to fulfill client contracts, while MCTS and ggMCTS fulfill client contracts in most cases. In all cases where client demand is not fulfilled, at least one client delivery did not meet client quality specifications due to feedstock and/or process variability. Over 100 simulations, greedy averaged 9.2 rejected deliveries per 14day simulation with each rejected delivery averaging 170 tonnes. MCTS averaged 5.0 rejected deliveries per 14-day simulation with each rejected delivery averaging 144 tonnes. Greedy-guided MCTS averaged 4.3 rejected deliveries per 14-day simulation with each rejected delivery averaging 144 tonnes. Table 6 – Profit relative to the MILP benchmark over 100 14-day simulations for Scenario B. Approach P10 (%) P50 (%) P90 (%) MCTS 94.9 105.6 124.1 ggMCTS 99.5 107.6 122.9 Table 7 – Average client contract fulfillment over 14-day simulations for each solver. Approach P10 (%) P50 (%) P90 (%) Greedy 78.0 89.8 96.4 MCTS 89.5 95.7 99.3 ggMCTS 88.5 96.1 99.1 4.3 Solver Decision-Making The blending network along with the flow assignments chosen by the MCTS solver averaged over the representative simulation (the 14-day simulation in which MCTS had median performance relative to the greedy benchmark) are plotted in Figure 5. The solution exhibits a simple yet effective blending pattern, with the thickness of edges proportional to the amount of material sent in that flow. Each node reports the average composition leaving that node. (The process node labels use shorthand to describe the processing pathways: “Wash” refers to washing only, “Float” refers to the addition of flotation, and “Grind” refers to the addition of grinding.)
RkJQdWJsaXNoZXIy MTM0Mzk2