The Parrot Optimizer (PO) is a new optimization algorithm based on the behaviors of trained Pyrrhura Molinae parrots. In this paper, an improved PO (IPO) is proposed for solving global optimization problems and training the multilayer perceptron. The basic PO is enhanced by using three improvements, which are aerial search strategy, modified staying behavior, and improved communicating behavior. The aerial search strategy is derived from Arctic Puffin Optimization and is employed to enhance the exploration ability of PO. The staying behavior and communicating behavior of PO are modified using random movement and roulette fitness–distance balance selection methods to achieve a better balance between exploration and exploitation. To evaluate the optimization performance of the proposed IPO, twelve CEC2022 test functions and five standard classification datasets are selected for the experimental tests. The results between IPO and the other six well-known optimization algorithms show that IPO has superior performance for solving complex global optimization problems. The results between IPO and the other six well-known optimization algorithms show that IPO has superior performance for solving complex global optimization problems. In addition, IPO has been applied to optimize a multilayer perceptron model for classifying the oral English teaching quality evaluation dataset. An MLP model with a 10-21-3 structure is constructed for the classification of evaluation outcomes. The results show that IPO-MLP outperforms other algorithms with the highest classification accuracy of 88.33%, which proves the effectiveness of the developed method.

Fire smoke detection plays a pivotal role in our life. For the optical fire smoke detector, it is easy to produce false alarms due to the misidentification of non-fire aerosols, such as dust and water mist. In this research, four multilayer perceptron (MLP) models are established and applied to discriminate the fire smokes and non-fire aerosols. A new hybrid algorithm (HBOBES) is proposed to optimize the parameters of MLP, which incorporates three search phases, including space selecting, promiscuous and restrictive mating, and consortship and extra-group mating, which are derived from the basic bonobo optimizer (BO) and bald eagle search algorithm (BES). Moreover, a adaptive tent chaos mapping technique is introduced into the first two phases to increase the population diversity. In the experiments, eight standard classification datasets and sixteen self-established aerosol classification datasets are applied to evaluate HBOBES’s optimization performance on training the MLP models. The results show that the proposed HBOBES ranks first overall, showing the merits in training MLP models compared to eight other algorithms. For the aerosol classification problem, it is found that both the optimized 3-7-2 and 3-7-3 MLP models have achieved the highest classification accuracy of approximately 95%. Black smokes can be identified with 100% classification accuracy. Therefore, this paper provides an effective and feasible approach to distinguish fire and non-fire aerosols by using the MLP classifier optimized by optimization algorithm, which has practical significance for the development of intelligent optical smoke detector.