Past literature shows that nitrous oxide (NOx) emission can be reduced by operating a spark ignition (SI) engine at the stoichiometric condition with high exhaust gas recirculation (EGR) levels. However, an engine, whose dynamics are typically unknown, will exhibit instability due to cyclic dispersion in heat release. A suite of novel neural network (NN) control schemes is developed to reduce the cyclic dispersion in heat release by using fuel as the control input. A separate control loop is designed for controlling EGR levels. The first NN scheme uses the total fuel and air as state variables for feedback control whereas a heat release-based output feedback scheme is developed next to relax the need for state variable measurements. The stability analysis of the closed loop system is demonstrated for both the schemes. No offline training phase is needed since online NN learning is utilised. Simulation and experimental results demonstrate that with control, the cyclic dispersion is reduced by 30%, NOx by 80% from stiochiometric levels and unburned hydrocarbons by 28% from the uncontrolled scenario.