-
Notifications
You must be signed in to change notification settings - Fork 0
Sub module: Search Space Control Method
This module explores the extent of the search space to supplement the inadequate prior knowledge of a bounded region that restricts the search of the mode choice model's extrema. The pseudo-code can be summarized as follows:
This module is composed of two sequential stages:
-
The first stage returns the nudges based on the following points:
- how much L1 norm for a mode split is reduced?
- which directionality (+ve or -ve) is required for such L1 norm reduction?
- magnitude of relative scaled (up/ down)
d_L1/d_intercept
intercept to achieve holistic L1 improvement?
-
The second stage returns the nudges based on the exponentially weighted average of the gradients and uses these gradients to update the intercept values. This method optimizes the mode choice analysis objective function reducing the oscillations (overshooting and diverging) that occur before reaching the global minimum. The parameter
beta = 0.9
provides friction to the acceleration termd_L1/d_intercept
(denoted asd_L1/d_m
) and the velocity termv_dintercept
(denoted asv_dm
) while the optimizer finds the path to the lowest L1 norm.
BEAMPyOpt
How to use
Mode Choice Analysis Model
Optimization Model
- Architecture
- Sub-module: Bayesian Optimization
- Sub-module: Search Space Control Method
- Memory Bank
- Parallelization and Multi-client Communication
- References
Evaluation
Contributing to BEAMPyOpt