Cartesian Genetic Programming (CGP) has often been applied for Neural Architecture Search (NAS). However, the performance of CGP is less than ideal when searching for architectures with limited computing resources. To better facilitate NAS with limited computing resources, this paper proposes a crossover operator, a light-weighted age mechanism, and two adaptive mutation operators as the novel components in our Efficient Cartesian Genetic Programming (ECGP) method. To assess the performance of ECGP, we conduct extensive experiments on three text classification task datasets. The experimental results demonstrate that ECGP outperforms other NAS methods, requiring only hundreds of fitness evaluations to find architectures with competitive accuracy compared with human-designed models. Additionally, the ECGP-evolved architectures are shown as converging fast and stably, and having high-level transferability with merely a 1-2% accuracy drop. Ablation studies demonstrate the effectiveness of the proposed operators and age mechanism, and identify GRU as the most critical function in the text classification task. Finally, we summarize three design principles observed from the ECGP-evolved architectures that are in line with human-design strategies. To the best of our knowledge, this work introduces the first attention-derived NAS benchmark for the text classification task.

Constraint-handling techniques and genetic operators are two crucial components in constrained multi-objective evolutionary algorithms (CMOEAs). Recent research in most of CMOEAs has primarily focused on adaptive designs of these components to address various constrained multi-objective optimization problems (CMOPs). However, the evolutionary process of solving a CMOP can involve various characteristics, such as continuity, discreteness, degeneracy, or some combination thereof, necessitating the tailored selection of constraint-handling techniques and genetic operators across different generations. This study conceptualizes these selections as a temporal sequence of constrained handling selection, where the time means the generation number. We argue that discovering the systematic patterns within the sequence based on the historical data of applying different selections significantly improves the performance of CMOEAs in finding Pareto optimal solutions. Based on this conceptualization, we propose a CMOEA with a deep reinforcement learning model for solving CMOPs. Specifically, the deep reinforcement learning model dynamically refines the selection of constraint-handling techniques and genetic operators for upcoming generations by learning from the performance of previous selections, thereby enhancing the predictive accuracy for subsequent selections. Experiments are conducted to validate the performance of the proposed algorithm against nine CMOEAs on thirty-seven benchmark problems and an unmanned aerial vehicle path planning problem. Experimental results show that the proposed algorithm substantially outperforms the compared algorithms regarding the obtained Pareto optimal solutions. Additionally, the results verify that discovering the systematic patterns within the sequence for CMOEAs has a positive impact on solving CMOPs in terms of objective optimization and constraint satisfaction.

Despite deep neural networks have made outstanding achievements in many static tasks, when faced with a continuous stream of data, they suffer from catastrophic forgetting since the previous data is usually inaccessible. Stored data or generative model is commonly used for maintaining the model performance but with memory utilization and privacy safety issues. Prototype-based methods address these issues by keeping only one prototype for each class but with limitations in its ability to trade-off the model stability and plasticity. In this paper, a novel exemplar-free class-incremental learning method is proposed which improves the stability of the representation learning and the decision boundary to a great degree. First, based on the results of our exploration into the impact of the batch normalization (BN) layer on representation learning, we propose to remove the BN layer (RBNL) in the incremental training phase to improve the stability of model representation learning. Then, to further maintain the feature space, we design the prototype mixing (PM), which expands the deep features by randomly and linearly combining prototypes of the old classes to generate hybrid prototypes with composite labels for fine-tuning the fully connected layer. Experimental results on three benchmark datasets, CIFAR-100, TinyImageNet, and ImageNet, show that our proposed method can effectively balance the stability and plasticity of the model, and outperforms the state-of-the-art works.