Merge pull request #211 from chhoumann/kb-299-future-works

[KB-299] Future works

report_thesis/src/sections/future_work.tex

@@ -1 +1,23 @@
-\section{Future Work}\label{sec:future_work}
+\section{Future Work}\label{sec:future_work}
+The findings of this study present several opportunities for future research.
+Firstly, regarding our data partitioning algorithm detailed in Section~\ref{subsubsec:dataset_partitioning}, we observed the significance of identifying the optimal percentile value $p$. 
+This value is crucial for minimizing extreme values in the test set while preserving its overall representativeness.
+Future work should explore quantitative methods for determining this optimal value.
+
+Another potential improvement to the validation and testing approach we delineate is incorporating supplementary extreme value testing after the preimary evaluation.
+This type of testing could be conducted using a small, separate subset of extreme values to assess the model's performance in these critical scenarios.
+For example, this might involve slightly reducing the percentile value $p$ and using the extreme values that fall within this reduced range to evaluate the model's effectiveness.
+
+Tackling the challenges of limited data availability has proven important, as we mention throughout our report.
+The small dataset size inherently restricts the number of extreme values present.
+These extreme values are crucial for enhancing the model's generalizability, as they represent the most challenging cases to predict.
+Future research could investigate methods for augmenting the dataset with synthetic data, including extreme values, to provide the model with more exposure to these cases during training.
+This is a hard task, as it requires the production of synthethic data for a physics-based process.
+We contemplate that some approximation may be sufficient, and could be used, for instance, as part of a transfer-learning project as inital training material.
+
+Future work should also consider further experimentation with the choices of base estimators and meta-learners.
+Our study demonstrated that various model and preprocessor configurations perform well.
+However, identifying the optimal configurations and meta-learner for a specific oxide remains a challenging task.
+In this study, we used a simple grouping method to ensure diversity in our base estimator selection, choosing from the top-performing configurations.
+We also experimented briefly with a grid search approach, which could be examined further.
+However, using a variation of the optimization framework we presented is likely to provide better trade-offs, as we have discussed.