| Advanced Traveler Information Systems (ATIS) and in-vehicle navigation systems emerged almost two decades ago as a solution for traffic congestion. However, these systems were not suited for daily use and could not adapt to routine activity patterns or learn travel behavior. Within this context different data collection tools were also developed to study not only travel behavior but also the adoption of this new technology. This research tries to address some of the deficiencies of these systems by proposing the design and implementation of a new data collection tool, called a travel simulator. This research will also use the data collected during the simulations to experiment with the ability of neural network models to learn the decision-making process that results in choices of activities, routes, and destinations, with the purpose of moving towards a new generation of ATIS.; The travel simulator is based on the activity approach and uses GIS as a developmental platform. The simulator has as a scenario a daily commute trip, with two intermediate activities and congestion taking place on the way to the first destination (traffic delay severities 20, 40, and 60 minutes are used). Subjects are asked to respond to this congestion by selecting a choice from a set of alternatives. The data analysis shows that changes in the time saved will modify the final choices made by subjects and result in shifts from one alternative to another as the severity of the delay increases.; Experiments were conducted to see if the decision-making process under the particular scenario outlined by the simulator could be learned and incorporated into data collection tools. A neural network (NN) model was used for this process, with two different topologies and two input datasets. The results do not completely favor a particular combination of topology or input, but are encouraging for continuing this vein of research. The models are able to predict more than 50% of the final choices for the sample of 41 cases. Additional experiments need to be conducted regarding the power of the neural network models, but this research indicates this should be a useful modeling approach. |
