International Journal of Computer Science and Artificial Intelligence Dec. 2012, Vol. 2 Iss. 4, PP. 33-39 - 33 - DOI: 10.5963/IJCSAI0204004 Optimizing Sign Placements for Crowd Evacuation on Road Network in Case of Tsunami Alert NguyenThi Ngoc Anh 1,2, a , Yann Chevaleyre 3, b , Jean Daniel Zucker 1,2,4, c 1 IRD, UMI 209, UMMISCO, IRD France Nord, 93143, Bondy, France 2 MSI, IFI, 42 Ta Quang Buu street, Hai Ba Trung District, Hanoi, Vietnam 3 Institut Galilee, University Paris-Nord 99, 93430, Villetaneuse, France 4 UPMC Univ Paris 06, UMI 209, UMMISCO, 75005, Paris, France a ngocanhfami@gmail.com; b yann.che@gmail.com; c jean-daniel.zucker@ird.fr Abstract- In recent years, the number of people affected by natural disasters and in particular tsunamis has been increasing. Artificial Intelligence and Operation Research approaches to simulate crowd evacuation and prepare cities for Tsunamis are of critical interest. Given an extremely simple model of human behavior, i.e. a memory-less stochastic agent, we address the problem of optimizing the placement of Tsunami evacuation signs with respect to evacuation time and casualties. Moreover, we formalise this optimisation problem as a Mixed Integer Linear Programming (MILP) problem and we ran some experiments with a MILP solver to give an early warning for the evacuation of the road network of Nhatrang city in Vietnam. Keywords- Agent Based Model;Optimization Sign Placement;Mixed Integer Linear Programming;Simulation INTRODUCTION Tsunamis have appeared frequently in recent years and have led to huge losses. On December 25, 2004 a tsunami in the Indian Ocean caused an earthquake of magnitude 9 on the Richter scale leading to approximately 230 000 deaths and on March 11, another tsunami in Tohoku, Japan caused an earthquake of magnitude 9 on the Richter scale leading to approximately 20 000 deaths [12] . Tsunamis lead to particularly severe damages when they occur near crowded residential areas or urban centres located on coastal zones. Since urban areas include complicated road networks, crowd evacuation on these networks requires identifying paths to safe locations [10, 7] . Although each urban area has its own peculiarities due to its topology, its infrastructure, its road network and even its cultural habit, the problem remains to escape to shelters as fast as possible. If a Tsunami alert is raised in the Philippines in the direction of Vietnam for example, the population of a coastal city like Nhatrang only has two hours to evacuate. In developing countries, the means to support evacuation include simple road signs since everyone cannot afford a smartphone with 3G capabilities that would support the reception of personalized evacuation routes as available in Japan for example. Furthermore, even in Japan, there are Tsunami route signs that support the evacuation and increase the awareness of Tsunamis. The problem of placing signs in cities to minimize casualties and evacuation time is thus critical. In this paper, we specifically address the problem of minimizing the average evacuation time by placing signs. This problem raises two research issues. The first issue is to accurately model a crowd and simulate it on a road network. As real road networks of many cities are now available in Geographic Information Systems (GIS), it becomes possible to simulate crowds over these networks of routes, and to assess the quality of a given sign placement in accordance. A number of crowd modeling approaches have been already applied separately or in combination to solve evacuation problems. Such approaches include cellular automata, lattice gas, social force, fluid dynamics, agent based approaches, game theory and experiments with animals [18, 8, 9, 10] . In these models, apart from Agent Based Models (ABM), the agents are always homogeneous and have similar behaviors. On the contrary, ABM considers each agent with its individual behavior and perception. Although ABM [4, 6, 16] are very flexible, they often require a lot more computation than their counterparts that may accommodate very large populations. The second issue is how to place signs and analyze the impact of such placements on the crowd behavior. This is mainly an optimization problem, which cannot usually be solved exactly. Earlier approaches involved in solving this optimization problem using heuristics, using the crowd simulator as a black-box to assess the performance of the candidate sign placements [18, 11] . This approach has two drawbacks: first, the simulator usually requires a lot of computation power, and the optimization process cannot thus explore the space of candidate solutions in depth. Also, even if we could explore sufficiently many solutions, nothing would guarantee us that we did not fall in a local minimum. In this paper, we tackle these two issues at once. By choosing a very simple model of crowds where the human behavior is modeled as memory-less stochastic agent, we can avoid the explicit use of a simulator. Instead, we formulate the optimization problem as a Mixed Integer Linear Program (MILP), in such a way that the dynamics of the agents are embedded inside this linear program. Then, the optimization problem can be solved using any state-of-the-art MILP solver. With this approach, we can either find optimal solution or get approximate solutions with some guarantees on the quality of the approximation, when applied to problems of larger scale. This of course is not possible with complex crowd models. The main motivation of this work is to bring an answer to the concrete problem of making Vietnamese coastal cities safer with respect to Tsunamis. Many American cities or Japanese cities are already considered Tsunami ready but much work remains to do for cities such as Nhatrang (a