A learning-based transportation oriented simulation system

1. A learning-based transportation oriented simulation system Theo A. Arentze, Harry J.P. Timmermans

2. Abstract • Albatross: activity-based model of activity-travel behavior, derived from theories of choice heuristics • The model predicts which activities are conducted when, where, for how long, with whom and the transport mode involved • Decision tree is proposed as a formalism to model the heuristic choice

3. Conceptual considerations regarding decision making and choice behavior • Postulate: activity participation, allocation and implementation fundamentally take place at the level of household • Decisions: • Long term: marital status, number of children, choice of work and workplace, purchase of transport mode • Short term • Decisions influence the generation of activity calendars

4. Constraints • Situational constraints: can’t be in two places at the same time • Institutional constraints: such as opening hours • Household constraints: such as bringing children to school • Spatial constraints: e.g. particular activities cannot be performed at particular locations • Time constraints: activities require some minimum duration • Spatial-temporal: constraints an individual cannot be at a particular location at the right time to conduct a particular activity

5. Choice behavior • Models used to rely on utility-maximization • Albatross assumes that choice behavior is based on rules that are formed and continuously adapted through learning while the individual is interacting with the environment (reinforcement learning) or communicating with others (social learning). • Exploration vs. exploitation dilemma

6. Learning theory sum • Albatross is based on a learning theory which implies that rules governing choice behavior are: • heuristic • context-dependent • adaptive in nature

7. The scheduling model • Components: • a model of the sequential decision making process • models to compute dynamic constraints on choice options • a set of decision trees representing choice behavior of individuals related to each step in the process model a-priori defined derived from observed choice behavior

8. Assumptions • Skeleton refers to the fixed and given part of the schedule • Flexible activities: optional activities added on the skeleton

9. The process model

10. The inference system • For each decision, the model evaluates dynamic constrains • The implementation of situational, household and temporal constraints is straightforward • We will look at space–time constraints and choice heuristics determining location choices

11. Feasible locations (step 6)

12. Heuristics • Having defined the location choice set, the proposed set of heuristics then define alternative ways of trading-off required travel time against attractiveness of locations.

13. Heuristics

14. Decision tree induction • Condition-action rules: • Albatross generates the best tree based on techniques from C4.5 (Quinlann, 1993), CART (Breiman et al., 1984) and CHAID (Kass, 1980)

15. Deriving decisions • Use a probabilistic assignment rule. The probability of selecting the q-th response for each new case assigned to the k-th node is:where fkq is the number of training cases of category q at leaf node k and Nk the total number of training cases at that node.

16. Testing the model

17. Results of inducing decision trees

18. Branch of time-of-day tree

19. Performance of Albatross • The eventual goodness-of-fit of the model can be assessed only by a comparison at the level of complete activity patterns • Eventual output of Albatross is trip matrices

20. Number of activities of predicted and observed patterns

21. Correlation coefficient

22. Summary • Use of decision trees for choice heuristics, resulting in a considerable, but varying improvement over a null model • A sample size of 2000 household-days suffices to develop a stable model • Transferability of the model to another context than in which it was developed remains to be studied • Model can be extended: use models of reinforcement learning (as opposed to supervised learning which are implicitly used by decision trees)