Task 2b: Work Locations (VALIDATION)

[Hussein-Mahfouz]

I'm currently using two approaches for assigning work locations (see work.py)

1 - select_work_zone_iterative() An iterative approach (iterating over individual activities and filling in a matrix that has constraints in each cell)

2 - select_work_zone_optimization(). Formulated as a linear optimization problem. Objective is minimize deviation of values from reference data. Two forms of deviation: (a) total deviation (sum of all deviations), and (b) maximum deviation among all ODs

Validation metrics

How do we test how well each approach performs? Our reference data (which is being used in both approaches) is the census commuting matrix

Some suggestions from this paper: A neural network approach for population synthesis

Error Based measures (compare degree of difference) 

• Total Absolute Error
• Root Mean Square Error (RMSE)
• Standardized RMSE
• Mean Absolute Error (MAE) 
• Population Error Rate

Similarity Measures (compare degree of similarity)

• Jaccard similarity coefficient
• Intersection rate
• Difference rate

I'm not sure how to use the similarity measures, as they are used for binary data. Should the values be converted to binary values based on threshold values (i.e. the predicted flow for an OD is considered the same as the actual flow if they are within x% of each other)?

Other approaches

There are probably better ways to approach this problem. Here are a couple of repos I've come across that use spatial interaction models to constrain a discrete OD matrix.

• https://github.com/YannisZa/ticodm (paper: https://onlinelibrary.wiley.com/doi/10.1002/sta4.656)
• https://github.com/YannisZa/GeNSIT
• https://github.com/ThGaskin/NeuralABM