For large applications with multiple ABs, the trade-off between speedup and error becomes complex. Often the optimum configuration of ALs in the various ABs are not obvious, especially if the approximation of internal ABs influences the number of iterations of the outer loop. Off-line exhaustive search can be possible only for a small number of approximate program configurations[ 43 ], and the majority of the previous approaches used various heuristic search strategies based on representative inputs[ 7 , 14 , 26 , 38 ], or dynamically tuned the ALs based on the observed errors from the approximated regions[ 7 , 25 , 38 , 44 ]. While many of these techniques identify and leverage properties of specific code patterns in ABs, they typically apply the same transformation for the entire execution and/or input. Such fixed optimization choices may lead to rigid transformed programs that miss fine-grained optimization opportunities