Still, Ipek et al.’s methodology has a key limitation: they do not propose a generalizable way to target an objective function (performance in their case). Because it is intuitive that bus utilization and throughput (and ultimately performance) correlate strongly for memory-intensive applications, it was natural and entirely appropriate for them to take a completely ad hoc approach to designing the RL reward function, by trivially rewarding load/store commands over precharge and activate commands. Unfortunately, this approach becomes much more difficult in other important scenarios that target more sophisticated objective functions (e.g., metrics that combine performance, energy, and/or fairness).