Input PPO explores whether a reinforcement-learning controller can improve the inputs used by another model. The experiment starts with Iris classification data and a small classifier. The controller supplies input representations, and the environment trains the classifier on them.
The feedback combines changes in training loss and classification accuracy. This makes the input representation part of the optimization problem, rather than treating the original features as fixed.
The current scope is a small supervised-learning experiment. It does not establish that the approach transfers to other datasets or larger models. A useful next comparison is whether any gain survives a held-out evaluation against the same classifier trained on the original features.