Multi-day activity models: An extension of the MultiAgent Transport Simulation (MATSim)

Abstract

MATSim is a large-scale multi-agent, activity-based transport simulation model. It can simulate the mobility of each person in an region, managing millions of agents in reasonable computation times. However, it is designed to simulate and reach a user equilibrium for a period of one day. This restricts the study of current transport planning challenges. Recent studies show that the behavioural variety of travellers can not be well analysed with only one day simulation results. Development of advanced time consumption travel models require observations of at least a week for calibration purposes, because a complete cycle of work and leisure must be included. To expand the standard MATSim time horizon, few changes in the current implementation have to be made. However, there are two reasons why the standard MATSim is not ideal for multi-day scenarios. First, the evolutionary algorithm of MATSim takes too long to reach a user equilibrium with longer periods, because of the combinatorial increase in the number possible activity chains to test per agent. The second reason why MATSim is not ideal for multi-day simulations is the difficulty of preparing multi-day plans to start the evolutionary process. In this paper a new approach for multi-day simulations is presented. To solve the mentioned MATSim drawbacks, a differentiation between fixed activities and flexible activities is proposed. An extension of MATSim to initialize agents with incomplete plans of fixed activities and schedule on-the-fly flexible activities is designed implemented and tested. Two case studies were prepared to evaluate the new approach. They show that the process is computationally feasible. The first iteration (where all agents plan flexible activities on-the-fly) takes 3.5 minutes for a 1% sample (37,425 agents) of the Singapore MATSim scenario. This is achieved running 20 threads for parallel flexible activities planning, and using 30GB of RAM memory. In the first case study, the new approach is compared with a standard MATSim evolutionary process. In 100 iterations, the average utility improves 100% more than the average utility increase with the current implementation. This happens because agents can schedule new flexible activities with new conditions of travel times every iteration. However, the evolutionary process needs 66% more RAM memory, and iterations are 47% slower than the standard MATSim. In the second case study a multi-day scenario with an unexpected event is simulated. Agents react on-the-fly to the event by planning different activities at different places and/or different times, while a few others make planning mistakes traveling to the affected area. Comparative analyses between the affected day and the other days are performed.