7 7.7. Results 167 indicate that the strategies of CBM, SchM, and RM are less e"cient for older pipes due to their higher failure probability. Regarding the mean pipe severity level to assess the impact of various policies on pipe deterioration, as shown in Table 7.6. Our analysis reveals a notable correlation between the average actions per policy, detailed in Table 7.5, and the mean pipe severity level. Specifically, the Agent-G control strategy tends to maintain pipes within a severity level of k ↓ [1,2,3], whereas the Agent-E, CBM, SchM, and RM policies often result in higher severity levels k ↓ [4,5, F], which correlates with increased policy costs. Table 7.6: Percentage (%) of severity level per policy obtained with Agent-E, Agent-G, CBM, SchM, and RM, evaluated over 100 episodes in the test environment for di!erent pipe ages. Pipe age Severity Agent-E Agent-G CBM SchM RM 0 k =1 59.77% 58.75% 59.94% 59.84% 58.88% k =2 33.27% 39.14% 32.67% 38.05% 33.15% k =3 5.39% 1.70% 6.00% 1.79% 6.36% k =4 1.38% 0.28% 1.13% 0.26% 1.30% k =5 0.18% 0.13% 0.25% 0.04% 0.31% k =F 0.01% 0.01% 0.01% 0.01% 0.01% 25 k =1 50.49% 41.72% 46.88% 39.07% 46.62% k =2 38.96% 55.27% 43.09% 55.55% 40.86% k =3 8.37% 2.63% 8.48% 4.85% 9.80% k =4 1.37% 0.29% 1.18% 0.41% 1.51% k =5 0.78% 0.07% 0.36% 0.10% 1.18% k =F 0.02% 0.01% 0.02% 0.01% 0.03% 50 k =1 57.93% 44.65% 55.01% 40.92% 54.36% k =2 32.58% 51.40% 36.14% 50.46% 33.09% k =3 7.50% 3.29% 7.20% 7.34% 9.32% k =4 1.31% 0.39% 1.19% 0.59% 1.64% k =5 0.65% 0.25% 0.43% 0.67% 1.55% k =F 0.03% 0.02% 0.02% 0.03% 0.03% To summarise, our findings indicate that the Agent-G’s policy, derived using DRL, implements a dynamic management strategy that varies with the pipe’s age. This strategy encompasses a more passive approach with new pipes, transitioning to active intervention as the pipes age. This indicates the agent’s preference for more frequent maintenance actions rather than allowing pipe failures, which incur higher penalties and replacement costs. Moreover, Agent-G outperforms Agent-E, illustrating the impact of the deterioration model assumption. Specifically, Agent-G’s prognostic model used during training aligns more closely with the test environment’s deterioration pattern than
RkJQdWJsaXNoZXIy MjY0ODMw