COMP 790

Learning "Internal goals" Parameter for Each Agent In a Department Store Environment

Goal of the Project

Realistic crowd simulation in grocery store based on shopping pattern. Predict a set of internal goals for each agents using machine learning.


Simulation in gaming, designing evacuation plans. For example, in a department store, placing emergency exit closer to a densed aisle probably will make any evacuation process faster.

State of the art

Social LSTM - Human Trajectory Prediction in Crowded Spaces uses LSTM model predicts trajectory which can be viewed as a sequence of a task.

Leveraging Long-Term Predictions and Online-Learning in Agent-based Multiple Person Tracking integrates agent-based crowd models and the high order particle filter to reason about pedestrians' motion and interaction and adaptively estimate their internal goals.

Regularity and Conformity: Location Prediction Using Heterogeneous Mobility Data integrates both the regularity and conformity of human mobility and their mutual reinforcement

Plan to accomplish over the semester

Realistic crowd distribution for each aisle/surrounding area in a store, instead of some random movements. It should be able to give a feel about what its like on a regular day at a department store.

Tentative Schedule

By the end of week 3

  • Data preperation.
  • Study machine learning models that are used to predict a sequence of task.
  • Add internal goals as parameters.

By the end of week 6

  • Implement machine learning model

By the end of week 9

  • Run simulation using learned parameter


Obtained store layout, grocery list and transaction time from dataset. Created a fictitious path by following- Starting with a random item from the list, found the next closest item, and continued to do it until all the items are found. Approximated customer density around different aisle in a given timeframe. For 50 different permutations, average density around each aisle: