Interestingly, the expectation of finding a bank closed in Ipek is significantly higher than in MORSE-P. It turns out that this is in part a potentially undesirable side effect of Ipek et al.’s imposition that a NoOp be allowable only if no legal commands (in particular, Precharge) exist, which was put in place to speed up convergence [17]. This means that Ipek is hardwired to closing pages (instead of doing nothing) when there are no other options available, regardless of the long-term benefit of doing so. In the case of MORSE-P, the scheduler learns by itself that it may ‘œbypass’ this restriction, by judiciously exploiting the PwDn/PwUp actions available in the DDR3 interface (i.e., it can force NoOp to be a legal choice indirectly by powering down a rank, effectively making Precharge to any bank within that rank ineligible until the rank is powered back up). As indicated before, the reward values obtained by our GA-based procedure hint at the potential benefit of this subterfuge, as it assigns a slightly positive aggregate reward to PwDn+PwUp, and a definitely positive reward to NoOp.