Recreation simulation modeling, when combined with intelligent monitoring, is becoming a valuable tool for natural resource managers. The goal of recreation simulation is to accurately model recreational use, both current and future. Models are applied to gain a thorough understanding of the characteristics of recreation. Indicator variables such as visitor experience, carrying capacity and impact on resources can be computed. If the model is valid it can be used to predict future use as well as to investigate the effect of new scenarios and management decisions. Recent research has focused on agent-based modeling techniques. Recreators are represented by autonomous, intelligent agents that travel across the landscape. A central issue is the model used for agent travel decisions. Current techniques range from replicating trips exactly to making local, intersection level decisions based on probability. But little attention has been paid to justifying these models. In this work we examine a range of probabilistic models. The models differ in the length of the Markov chain used to compute agent decisions. The length of the chain ranges from zero (local decisions only) to infinity (exact trip replication). We test the length of the chain on held out data for validation. We show that the choice of model strongly influences the validity and results of the simulation. To test these models we present a framework for automatically constructing agent-based models from an input set of GPS tracklogs. The GPS tracklogs are collected by volunteers as they recreate in natural areas. Traditionally, data on where recreators travel is collected in the form of trip diaries, filled out on paper by visitors or by interview. Other demographic and attitudinal data is also collected along with the actual route traveled. Although the additional information is valuable, the data must be collected and entered by hand. Paper diaries also place a significant time burden on visitors, reducing the compliance rate as well as skewing the results (ensuring only visitors with excess time participate). Using GPS devices to record visitor trips helps alleviate these problems. The framework for processing GPS trips and automatically building a model presented in this work significantly reduces the time required to build a model, lowers the cost and widens the applicability of recreation simulation modeling to new areas. GPS devices automatically record their data, requiring only that visitors turn the unit on and carry it with a marginal view of the sky. GPS use is also becoming more widespread among recreators. As more recreators use GPS to record their trips, data useful to modeling is becoming increasingly available. The steps in GPS driven model generation are as follows. First, the set of GPS tracklogs is combined to form the underlying travel network along which agents will travel. Each GPS tracklog is then traced along the network in order to determine what choices were made as the recreator traveled across the network. This produces a list of trip itineraries. Model parameters (probability tables) can then be computed from the trips. The length of the Markov chain used in the probability tables is a parameter to the model. The optimal value is found by testing the likelihood of heldout data for different chain lengths. This step is done automatically. Once the optimal length of the chain is chosen the model is complete and agent-based simulation can proceed. The entire framework for automatically producing GPS driven agent-based models is implemented in our TopoFusion GPS mapping software. We present results from two collections of GPS tracklogs from different trail systems. The first is from Tucson Mountain Park and is the result of a volunteer collection effort by the authors. A trails master plan is underway at the park, with input from our model. The second is a collection of tracks from mountain bike rides in the Finger Rock Wash area, collected by the author. Testing by held-out data on both GPS datasets indicates that current modeling methods are insufficient to model recreator travel decisions. The middle ground (neither exact replication nor local decisions) consistently performs better.