Observation function

[micheleurbani]

Observation function

An observation function is an Ecole environment that returns features of the actions in the action set.

The set of features depends on the problem representation we want to use.

Direct problem representation

Use problem attributes directly as in the canonical form {min cx, Ax <= b}.
Variable features are harder to characterize according to the problem under study. An example of generic features is that provided by Khalil.

Indirect problem representation

A solution to the problem is described as a partition of the set of tasks, where each element corresponds to the set of tasks assigned to a station.

Features are easier to figure out than with the direct representation.

Information extraction from ecole.scip.Model

We need a fast and reliable way to extract the indirect solution representation from the ecole.scip.Model (direct representation). The availability of problem information would help the extraction of variable values. For example, by providing the number of activities and the type of problem, it is possible to understand which variables have been used and how they are indexed, hence retrieving their value and computing, e.g., the list of unassigned tasks or the list of empty stations.

The problem can be divided into two subproblems: the first concerns retrieving information about the problem and it is tackled in the issue #4, whereas accessing information from the ecole.scip.Model concerns accessing the model at runtime and is tackled in the following.

Depending on the programming language, I see two possible ways to access information from ecole.scip.Model.

Python observation function

Exploiting the model name to retrieve the type of the problem, it is possible to access variables correctly due to the different MILP formulations that were adopted.

C++ observation function