Methods

[vivekaxl]

ePAL

Epsilon is a variable in ePAL that can be used to determine how close the points are to the actual Pareto Frontier. This variable is used to balance between quality of the solutions and the number of evaluations. For example, in wc-sol-3d-c4 - ePAL-0.01 has lower scores in Generational distance but higher evaluations where as ePAL-0.3 has higher scores in Generational Distance but lower evaluations.

MMRE-Prog

This is Guo's method, where we build an accurate model (based on threshold mmre scores) for each objective. Once the model is built, it is used to predict the dependent values (performance scores). A non-domination sort of these points (based on predicted values) yield the predicted pareto frontier.

Rank-Prog

This is based on our FSE submission, where the model is based on rank-difference for each objective. Once the model is built, it is used to predict the dependent values (performance scores). A non-domination sort of these points (based on predicted values) yield the predicted pareto frontier.

AL1

In this method,

1. Initial points are selected randomly (20 in our case)
2. Build a model for each objective
3. Model built in step 2, is used to predict the non-dominated points.
4. 10 samples (selected randomly) is then evaluated and moved to the training set
5. Repeat step 2-4 till stopping criterion is reached

Stopping criterion used in this case is: no new non dominated point is found. (lives = 3)

AL2 (preferred method)

In this method,

1. Initial points are selected randomly (20 in our case)
2. Build a model for each objective
3. Model built in step 2, is used to predict the non-dominated points.
4. cdom is used to score each point based on how many points (from step 3) it dominates.
5. The point which dominates most number of points is added to the training set. There are cases where the domination scores of all the elements are equal. In such cases a random point is added.
6. Repeat step 2-4 till stopping criterion is reached

Stopping criterion used in this case is : no new point is added to the existing non-dominated solution. (lives = 10)

[timm]

Q1: what are these models? "wc-sol-3d-c4"? a cloud sim config?

• need tables on the conclusions found by these methods

Q2: i think i missed something. what is epal?

[vivekaxl]

Q1: what are these models? "wc-sol-3d-c4"? a cloud sim config?

These are the datasets used in our FSE'17 paper.

Q2: i think i missed something. what is epal?

This is the Zulauga's paper from 2016. ePAL = epsilon Pareto Active Learning.