Fig. 2

Simple experiment demonstrating adaptable learning behavior starting with the default REINVENT 4 agent. 500 epochs of RL are run with a scoring function that rewards molecular weight \(\ge 1500\) Da, before it is switched in a second stage that rewards molecular weight \(\le 1500\) Da, showing the score a, molecular weight b, agent and prior likelihoods c and loss function d averaged over all molecules at the end of each epoch. The loss lower bound (Eq. 7) is also shown in d. A dashed line indicates the change of scoring function. The run used default settings: batch size of 128 and \(\sigma =128\)