GitLab

Introduces a new potential specialization mechanism: regardless of the current task, an agent will adapt its knowledge only when winner is better on his favorite task.

2) In SelectiveAdaptationSpecializationExperiment, depending on "deficiencyBased" the ranking may be done based on the reward or based on the agent's suitability with respect to the irrelevant features.

An agent can now rank tasks depending on how many of its the features it cannot perceive correctly are not part of the task's irrelevant features. The higher this number, the highest its unsuitability for performing this task.

The logic of numberOfRelevantFeatures is completely inversed. No everything is implemented based on numberOfIrrelevantFeatures.

2) Training examples are now task dependent when numberOfRelevantFeatures is greater than 0.
3) For the task's non relevantFeatures the value is always false, constituting them as constants.

1) Each task now has "relevantFeatures" which are saved and retrieved from File or generated randomly in different MultitaskEnvironment configurations.

2) The numberOfFeatures is now passed as a property at all multitasking experiments.