Electroencephalography (EEG) based emotion recognition is a widely preferred technique due to its noninvasiveness. Also, frontal region-specific EEG signals have been associated with emotional processing. Feature reductionbased optimized machine learning methods can improve the automated analysis of frontal EEG signals. In this work, an attempt is made to classify emotional states using entropybased features and Bayesian optimized random forest. For this, the EEG signals of prefrontal and frontal regions (Fp1, Fp2, Fz, F3, and F4) are obtained from an online public database. The signals are decomposed into five frequency bands, namely delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (14-30 Hz), and gamma (30-45 Hz). Three entropy features, namely Dispersion Entropy (DE), Sample Entropy (SE), and Permutation Entropy (PE), are extracted and are dimensionally reduced using Principal Component Analysis (PCA). Further, the reduced features are applied to the Bayesian optimized random forest for the classification. The results show that the DE in the gamma band and SE in the alpha band exhibit a statistically significant (p < 0.05) difference for classifying arousal and valence emotional states. The selected features from PCA yield an F-measure of 73.24% for arousal and 46.98% for valence emotional states. Further, the combination of all features yields a higher F-measure of 48.13% for valence emotional states. The proposed method is capable of handling multicomponent variations of frontal region-specific EEG signals. Particularly the combination of selected features could be useful to characterize arousal and valence emotional states. © 2021 by Walter de Gruyter Berlin/Boston.