Fig. 2: Workflow of DyRAMO (Dynamic Reliability Adjustment for Multi-objective Optimization).
From: A data-driven generative strategy to avoid reward hacking in multi-objective molecular design
DyRAMO explores an appropriate combination of the reliability level for each target property prediction. The exploration is conducted by repeating the following three steps. In step 1, a reliability level is selected for each property prediction, and the applicability domains (ADs) of the prediction models for the properties are defined based on the selected reliability levels. In step 2, molecular design is performed with the aim of satisfying the reliability levels given in step 1 for all property predictions (Aim 1) and achieving multi-objective optimization (Aim 2). This design process is conducted to obtain molecules that fall within all defined ADs and have optimized target properties. In step 3, the evaluation of the molecular design is performed from two aspects: the desirability of reliability levels selected in step 1 and the degree of optimization of predicted properties of the molecules designed in step 2. The DSS (Degree of Simultaneous Satisfaction of prediction reliability and multiple property optimizations) score was introduced to make this assessment. By repeating this cycle, an appropriate combination of the reliability levels is searched for so as to improve the DSS score. In DyRAMO, Bayesian optimization (BO) is introduced to accelerate this exploration.
