24406 research outputs found
Sort by
Optimized Long Short-Term Memory With An Adaptive Aquila Optimizer For Long-Term Dependency Problems
No full textRecurrent Neural Networks, particularly Long Short-Term Memory (LSTM) networks,
are widely applied in sequential data processing due to their ability to model temporal
dependencies. However, LSTMs still struggle to capture long-term dependencies due
to vanishing and exploding gradient issues, suboptimal hyperparameter tuning, and
architectural adaptability. These limitations reduce their effectiveness in dynamic and
real-world applications requiring high accuracy over long sequences. Motivated by
these challenges, this thesis introduces the Adaptive Aquila Optimizer (AAO), a novel
metaheuristic enhanced with a Sigmoid Factor to dynamically balance exploration and
exploitation. The AAO automates feature selection and hyperparameter tuning across
seven critical parameters: number of LSTM units, training epochs, batch size, dropout
rate, optimizer, learning rate, and activation function. A bidirectional LSTM
(BiLSTM) architecture, integrated with the AAO and visualized via the TensorBoard
dashboard, is proposed to improve long-term dependency modeling and learning
efficiency. Validation on three well-known classification datasets (Electrical Grid
Stability, Accelerometer Gyro Mobile Phone, and AI4I 2020 Predictive Maintenance)
demonstrates significant performance improvements of the proposed AAO-BiLSTM
model over traditional and state-of-the-art approaches. Achieving high accuracy,
precision, recall, and F1-score rates of up to 99.55%. The main contribution of this
work is the development of an optimized AAO-BiLSTM framework, which provides
an effective solution for mitigating vanishing gradients and improving long-term
dependency modeling in LSTM networks. This thesis advances LSTM-based systems
by introducing a robust, adaptive optimization strategy suitable for a range of
real-world time-series applications
