Electrical status epilepticus during sleep (ESES) is a serious condition that causes notable cognitive decline. It is characterized by distinct spike and slow-wave patterns on electroencephalograms (EEG). Clinical ESES diagnosis is extremely time-consuming and labor-intensive as it demands clinicians to manually interpret and count EEG screens. Existing automated diagnosis algorithms for ESES have major flaws, like struggling to adapt to complex spike-and-wave patterns and not fully exploiting the rich multi-view features of EEG. To overcome these issues, we propose a multi-view feature selection framework integrating reinforcement learning and attention mechanisms for automated ESES diagnosis. A CLEAN reward mechanism is introduced to address complex multi-objective feature selection challenges. Experiments on the clinical data consisting of 36 epilepsy patients prove the proposed method's remarkable spike-and-wave identification ability and high agreement with expert diagnoses. Our approach represents a significant step toward developing automated bedside ESES clinical diagnostic systems.