IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 430 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 SIMULATING TSUNAMI EVACUATION WITH MULTI-AGENTS AND DETERMINING A COUNTERMEASURE WITH AHP Kazuhira Kohara Chiba Institute of Technology Japan k-kohara@jcom.zaq.ne.jp Takuya Sugiyama Chiba Institute of Technology Japan kohara.kazuhiro@p.chibakoudai.ac.jp ABSTRACT We propose an integration method that uses agent-based modeling to simulate tsunami evacuation and the Analytic Hierarchy Process (AHP) to make a decision on a countermeasure. First, we created multiagent coast models that include a tsunami agent, shelter agents, road agents and evacuee agents. Second, we divided the coast into several districts and predicted the tsunami evacuation success/failure number of each district by using a computer simulation with multiagent coast models. Third, we considered several countermeasures (adding a shelter, adding an evacuation route) using that prediction. Fourth, we estimated the effects of each countermeasure. Finally, we use AHP to determine the best countermeasure against a tsunami disaster. Keywords: agent-based modeling; tsunami disaster; countermeasure decision making 1. Introduction Multiagent-based social simulations have been extensively investigated, and various attempts have been made to apply them to the layout design of supermarkets, stock markets, sales prediction and tsunami evacuation (Yamane et al. 2012; Panayi et al. 2012; Kohara et al. 2014; Saito et al. 2005). The Analytic Hierarchy Process (AHP) has been widely used for economic, political, social and corporate decision making (Saaty, 1980; Saaty et al. 1994; Saaty, 2001; Ginda et al. 2016). Previously, we investigated the real-world problem of predicting sales for stores and using that prediction to determine where to locate a new store (Kohara et al. 2014). We proposed an integrated method that uses agent-based modeling and the AHP to predict sales and to choose a new store location. First, we created multiagent town models that included store agents and consumer agents. We then estimated the predicted sales for each store by using a computer simulation based on multiagent town models. Finally, we used AHP to determine the location of a new store. mailto:k-kohara@jcom.zaq.ne.jp mailto:kohara.kazuhiro@p.chibakoudai.ac.jp IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 431 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 In this paper, we investigate another real-world problem of predicting the tsunami evacuation success/failure number and using that prediction to determine countermeasures against a tsunami disaster. We propose an integrated method that uses agent-based modeling and the AHP to predict the evacuation success/failure number and determine the countermeasure. First, we created multiagent coast models that consist of a tsunami agent, shelter agents, road agents and evacuee agents. Then, we estimated the predicted evacuation success/failure number by using a computer simulation with multiagent coast models. Finally, we used AHP to decide the best countermeasure against a tsunami disaster. The main features of our method are as follows: (1) We introduced a tsunami into the coast models as an agent. (2) We divided the coast into several districts, predicted the tsunami evacuation success/failure number of each district and considered several countermeasures (adding a shelter, adding an evacuation route) using that prediction. (3) We estimated the effects of each countermeasure. (4) We determined the countermeasure using the results of a multiagent simulation and AHP. 2. Multiagent coast models We created a multiagent coast model based on a popular coast in Shizuoka Prefecture where it is predicted that a large earthquake will occur in the near future. The width of the coast is 2 km. Since the size of the coast model is 100 cells by 200 cells, one cell corresponds to 10 m in each direction. Evacuees move one cell per step and 100 m per minute, so a minute corresponds to 10 steps. Evacuees move one cell per two steps on a sloping road. The tsunami moves two cells per step. Here, we assumed that a 10 meter high tsunami arrives 10 minutes after the earthquake, and that 60% of evacuees start to evacuate immediately, 30% of evacuees start at 5 minutes after the earthquake, and 10% of evacuees start at the time of the arrival of the tsunami. These assumptions are based on results from a questionnaire carried out after the large earthquake that occurred in Japan on March 11, 2011. There are four kinds of agents: a tsunami agent, shelter agents, road agents and evacuee agents. The number of evacuee agents is 3000, based on the published number of people bathing at the coast. The number of shelters is 10, based on the actual information. Evacuee agents move to a higher location in the same way as in the related work (Saito et al., 2005). When evacuees arrive at the intersection, they move according to traffic signs. If there is a shelter, they go to the shelter. Otherwise, they move to a higher location. Figures 1 and 2 show our multiagent coast model. In Figure 1, green lines show flat roads, orange lines show sloping roads and light blue shows the sea. IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 432 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Figure 1. Multiagent coast model In Figure 2, dark blue shows tsunami agents. The tsunami moves two cells per step. Figure 2. Multiagent coast model 3. Predicting evacuation success/failure number First, we divided the coast into five districts (A, B, C, D and E) as shown in Figure 3. We generated 600 people in each district and estimated the evacuation success number of each district. We examined an average number of 50 trials. Table 1 shows the results. The evacuation success number of district C is comparatively small. IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 433 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Figure 3. Dividing the coast into five districts Table 1 Estimated evacuation success number of each district Districts Estimated evacuation success number District A 472 District B 475 District C 452 District D 480 District E 483 Second, we generated 3,000 people on the coast and estimated the evacuation success number. Again, the number of shelters is ten. We assumed that a 10 meter high tsunami arrives at 10 minutes after the earthquake and leaves at 15 minutes after the earthquake. We also examined an average number of 50 trials. Table 2 shows the results. The number of evacuees who succeeded in reaching shelters was 1,397 and number of evacuees who succeeded in reaching high places was 1,076. Therefore, the total number of successes was 2,473 and total number of failures was 527. Table 2 Results of tsunami evacuation simulation of the current state Estimated number Success number to reach shelters 1,397 Success number to reach high places 1,076 Total success number 2,473 Total failure number 527 Third, we added a shelter (shelter K in district B, shelter L in district C, or shelter M in district D) or an evacuation route (route X in district A, route Y in district B, or route Z in district C) based on the above results and the actual map, as shown in Figures 4 and 5. IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 434 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Figure 4. Adding a shelter Figure 5. Adding an evacuation route Table 3 shows the results when a shelter is added. The average failure number for the 50 trials was 449 when adding shelter K, 372 when adding shelter L, and 343 when adding shelter M. Adding shelter M was the most effective, and adding shelter L was the second most effective. Table 3 Results of tsunami evacuation simulation in case of adding a shelter Success to shelters Success to high places Total number of success Total number of failure Current state 1,397 1,076 2,473 527 Adding shelter K 1,842 709 2,551 449 Adding shelter L 1,962 667 2,628 372 Adding shelter M 2,024 633 2,657 343 Table 4 shows the results when an evacuation route was added. The average failure number for the 50 trials was 236 when adding route X, 230 when adding route Y, and 164 when adding route Z. Adding route Z was most the effective and adding route Y was the second most effective. IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 435 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Table 4 Results of tsunami evacuation simulation in case of adding an evacuation route Success to shelters Success to high places Total number of success Total number of failure Current state 1,397 1,076 2,473 527 Adding route X 2,075 690 2,764 236 Adding route Y 1,874 897 2,770 230 Adding route Z 2,392 443 2,836 164 4. Determining a countermeasure by using AHP Figure 6 shows the relative measurement AHP model created for the task of deciding a tsunami evacuation measure. Here, we used the following four criteria: feasibility, evacuation success rate, cost, and time required to realize the countermeasure. In feasibility, a countermeasure whose feasibility is high is important. In success rate, a countermeasure whose success rate is high is important. In cost, a countermeasure whose cost is low is important. In required time, a countermeasure whose required time is short is important. Figure 6. Determining a countermeasure with AHP Here, we used the following four alternatives: (1) adding shelter L, (2) adding shelter M, (3) adding evacuation route Y, and (4) adding evacuation route Z. Table 5 shows four alternatives for a countermeasure against a tsunami disaster. Table 5 Four alternatives for a countermeasure against tsunami disaster Success to shelters Success to high places Total number of success Total number of failure Adding shelter L 1,962 667 2,628 372 Adding shelter M 2,024 633 2,657 343 Adding route Y 1,874 897 2,770 230 Adding route Z 2,392 443 2,836 164 IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 436 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Table 6 shows pairwise comparisons of four criteria when feasibility and success rate are most important. In this case, the weights of feasibility and success rate are the highest (their weights = 0.342). Consistency index means whether a pair comparison matrix is consistent or not. When the index is lower than 0.10, we judge that the pair matrix is consistent. Here, the consistency index is 0.041 and the pairwise comparisons are consistent. Table 6 Pairwise comparisons of four criteria when feasibility and success rate are most important Feasibility Success rate Measures cost Required time Weight Feasibility 1 1 2 3 0.342 Success rate 1 1 2 3 0.342 Measures Cost 1/2 1/2 1 4 0.226 Required time 1/3 1/3 1/4 1 0.091 Consistency index = 0.041 Table 7 shows pairwise comparisons of alternatives with respect to predicted feasibility. We will use existing hotels for shelters L and M, and construct new evacuation routes for Y and Z. Therefore, the feasibility of shelters L and M is higher than that of routes Y and Z. Route Z is shorter than route Y; therefore, the feasibility of route Z is higher than that of route Y. The weights of shelters L and M were the highest. Table 7 Pairwise comparisons of alternatives with respect to feasibility Shelter L Shelter M Route Y Route Z Weight Shelter L 1 1 6 2 0.368 Shelter M 1 1 6 2 0.368 Route Y 1/6 1/6 1 1/5 0.054 Route Z 1/2 1/2 5 1 0.211 Consistency index = 0.011 Table 8 shows pairwise comparisons with respect to success rate. The success rate is 0.945 (2836/3000) when adding route Z, 0.923 (2770/3000) when adding route Y, 0.886 (2657/3000) when adding shelter M, and 0.876 (2628/3000) for shelter L. The weight of shelter Z is the highest. Table 8 Pairwise comparisons of alternatives with respect to success rate Shelter L Shelter M Route Y Route Z Weight Shelter L 1 1/2 1/4 1/5 0.078 Shelter M 2 1 1/3 1/4 0.125 Route Y 4 3 1 1/2 0.306 Route Z 5 4 2 1 0.492 Consistency index = 0.016 IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 437 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Table 9 shows pairwise comparisons with respect to the measure cost. We will use existing hotels for shelters L and M; therefore, the measure cost is comparatively low. Shelter M is smaller than shelter L. Route Z is shorter than route Y. The weight of shelter M is the highest. Table 9 Pairwise comparisons of alternatives with respect to measures cost Shelter L Shelter M Route Y Route Z Weight Shelter L 1 1 8 3 0.317 Shelter M 1 1 9 4 0.499 Route Y 1/8 1/9 1 1/6 0.043 Route Z 1/3 1/4 6 1 0.146 Consistency index = 0.012 Table 10 shows pairwise comparisons with respect to required time. As route Z is short, required time to construct route Z is short. As shelter L is larger than shelter M, comparatively many rooms can be used for evacuees immediately. The weight of route Z is the highest. Table 10 Pairwise comparisons of alternatives with respect to required time Shelter L Shelter M Route Y Route Z Weight Shelter L 1 2 5 1/2 0.289 Shelter M 1/2 1 4 1/3 0.176 Route Y 1/5 1/4 1 1/6 0.059 Route Z 2 3 6 1 0.476 Consistency index = 0.022 Table 11 shows the final results when feasibility and success rate are most important. In this case, the weight of adding route Z is highest because the weights of adding route Z with respect to feasibility and success rate are comparatively high. Table 11 Final results of AHP when feasibility and success rate are most important Alternatives Results Adding shelter L 0.250 Adding shelter M 0.297 Adding route Y 0.138 Adding route Z 0.317 5. Additional studies 5.1 Additional study on changing tsunami height The tsunami height is based on the expected information which is between 5 and 10 meters. Table 12 shows an additional study on tsunami height. First, we assumed the IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 438 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 tsunami height is 10 meters. Then, we changed the tsunami height to 5 meters. The evacuation failure number was 169. The tsunami height is crucial for tsunami evacuation; however, a countermeasure against a 10 meter high tsunami is important. Table 12 Additional study: changing tsunami height Tsunami height Success to shelters Success to high places Total number of success Total number of failure 5 meters 1,420 1,411 2,831 169 10 meters 1,397 1,076 2,473 527 5.2 Additional study on changing tsunami arrival time The tsunami arrival time is based on actual information. Tsunamis have arrived at 10 minutes after the earthquake on average over the past hundred years. Table 13 shows an additional study on tsunami arrival time. First, we assumed the tsunami arrives 10 minutes after the earthquake. Then, we changed tsunami arrival time to 20 minutes or 30 minutes after the earthquake. The evacuation failure number was 508 for 20 minutes or 461 for 30 minutes. Anyway, evacuate immediately! Table 13 Additional study: changing tsunami arrival time Tsunami arrival time Success to shelters Success to high places Total number of success Total number of failure 10 minutes 1,397 1,076 2,473 527 20 minutes 1,860 632 2,492 508 30 minutes 1,903 636 2,539 461 5.3 Additional study on changing tsunami speed Table 14 shows an additional study on tsunami speed. First, we assumed the tsunami moves 2 times faster than people’s speed. Then, we changed it so that the tsunami moves 4 times faster than people’s speed. The evacuation failure number was 922 when the tsunami moves 4 times faster than people’s speed. Anyway, evacuate immediately! Table 14 Additional study: changing tsunami speed Tsunami speed Success to shelters Success to high places Total number of success Total number of failure 2 times faster 1,397 1,076 2,473 527 4 times faster 1,347 731 2,078 922 5.4 Additional study on changing percentage of evacuation consciousness Table 15 shows an additional study on changing the percentage of evacuation consciousness. First, we assumed that 60% of evacuees start to evacuate immediately, 30% of evacuees start at 5 minutes after the earthquake, and 10% of evacuees start at the time of arrival of the tsunami, based on questionnaire results. Then, we changed the IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 439 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 percentage of evacuation consciousness into 70%:25%:5% and 70%:29%:1%. The evacuation failure number was 296 for 70%:25%:5% and 123 for 70%:29%:1%. Again, evacuate immediately! Table 15 Additional study: changing percentage of evacuation consciousness Percentage Success to shelters Success to high places Total number of success Total number of failure 60%:30%:10% 1,397 1,076 2,473 527 70%:25%:5% 1,584 1,120 2,704 296 70%:29%:1% 1,571 1,306 2,877 123 5.5 Additional study on changing pairwise comparisons of four criteria Table 16 shows an additional study on changing pairwise comparisons of the four criteria when required time is most important. In this case, the weight of required time is 0.549. Table 16 Pairwise comparisons of four criteria when required time is most important Feasibility Success rate Measures cost Required time Weight Feasibility 1 2 4 1/3 0.239 Success rate 1/2 1 3 1/4 0.147 Measures Cost 1/4 1/3 1 1/6 0.067 Required time 3 4 6 1 0.549 Consistency index = 0.020 Table 17 shows the final results when required time is the most important. In this case, the weight of route Z is highest because the weights of route Z with respect to the success rate and required time are comparatively high. Table 17 Final results of AHP when required time is most important Alternatives Results Adding shelter L 0.279 Adding shelter M 0.236 Adding route Y 0.093 Adding route Z 0.394 Table 18 shows another additional study on changing pairwise comparisons of the four criteria when cost is most important. In this case, the weight of cost is 0.495. IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 440 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 Table 18 Pairwise comparisons of four criteria when measures cost is most important Feasibility Success rate Measures cost Required time Weight Feasibility 1 3 1/2 5 0.310 Success rate 1/3 1 1/4 3 0.134 Measures Cost 2 4 1 6 0.495 Required time 1/5 1/3 1/6 1 0.061 Consistency index = 0.026 Table 19 shows the final results when cost is most important. In this case, the weight of shelter M is highest because the weights of shelter M with respect to feasibility and measures cost are comparatively high. Table 19 Final results of AHP when measures cost is most important Alternatives Results Adding shelter L 0.299 Adding shelter M 0.389 Adding route Y 0.083 Adding route Z 0.233 6. Conclusion We proposed integrating agent-based modeling with the AHP for predicting a tsunami evacuation success/failure number and making decisions about countermeasures against a tsunami disaster. First, we created multiagent coast models that include a tsunami agent, shelter agents, road agents and evacuee agents. Second, we divided the coast into five districts and estimated the evacuation success/failure number of each district by using a computer simulation with multiagent coast models. Third, we added a shelter or an evacuation route and estimated the failure number. Finally, we applied the AHP with four criteria. We applied our method to an actual coast and showed its effectiveness. We also reported additional studies on changing tsunami height, changing tsunami arrival time, changing tsunami speed, changing percentage of evacuation consciousness, and changing pairwise comparisons of four criteria. In future work, we will apply our method to other cases, other coasts and other types of AHP and ANP for decision making (Saaty, 1996; Saaty et al. 2013). IJAHPArticle: Kohara, Sugiyama/Simulating tsunami evacuation with multi-agents and determining a countermeasure with AHP International Journal of the Analytic Hierarchy Process 441 Vol. 8 Issue 3 2016 ISSN 1936-6744 http://dx.doi.org/10.13033/ijahp.v8i3.406 REFERENCES Ginda, G., & Dytczak, M. (2016). A survey of AHP and ANP applications in civil engineering and urban management. ISAHP 2016. Kohara, K., & Sekigawa, D. (2014). Sales prediction with multiagent town models and deciding stores location with AHP. ISAHP 2014. Panayi, E., et al. (2012). Agent-based modeling of stock markets using existing order book data. MABS’12 (13th International Workshop on Multi-Agent Based Simulation). Doi: 10.1007/978-3-642-38859-0_8 Saaty, T. (1980). The Analytic Hierarchy Process. New York: McGraw-Hill. Doi: http://dx.doi.org/10.1080/00137918308956077 Saaty, T., & Vargas, L. G. (1994). Decision making in economic, political, social and technological environments with the Analytic Hierarchy Process. Pittsburgh: RWS Publications. Doi:10.1016/0898-1221(95)90109-4 Saaty, T. (1996). The Analytic Network Process. Arlington: Expert Choice. Saaty, T. (2001). Decision making for leaders: The Analytic Hierarchy Process for decisions in a complex world, Pittsburgh: RWS Publications. Saaty, T., & Vargas, L. G. (2013). Decision making with the Analytic Network Process: Economic, political, social and technological applications with benefits, opportunities, costs and risks (Second Edition). Springer Verlag. Doi: 10.1007/978-1-4614-7279-7_1 Saito, T., & Kagami, H. (2005). Simulation of evacuation behavior from tsunami utilizing multiagent system. Journal of Architecture and Planning, Transactions of AIJ, 597, 229- 234. (in Japanese) Doi: http://doi.org/10.3130/aija.74.627 Yamane, S. et al. (2012). Agent-based social simulation for a checkout layout design of a specific supermarket. MABS’12 (13th International Workshop on Multi-Agent Based Simulation). http://dx.doi.org/10.1080/00137918308956077 http://dx.doi.org/10.1016/0898-1221(95)90109-4 http://doi.org/10.3130/aija.74.627