For training a deep learning model, one has to get the parameters necessary
to satisfy an objective function. In general, the objective is to minimize the loss
encountered throughout the learning process. A type of supervised learning involves
feeding data samples and their matching outputs to a model. A framework works to get
the generated output nearer to the desired result by comparing its results to the planned
outcome and taking into account any variations. Optimization algorithms have been
utilized to do this. Multiple cycles of optimization are carried out till the process is
completed in order to increase the accuracy of the model. Numerous optimization
techniques have been developed to address the challenges associated with the learning
process. The techniques addressed involve Adadelta, Adagrad, Adam, and stochastic
gradient descent. With Adam and Adagrad, Pond1 has a desirable training result of
0.96 at epoch 200, and Pond2 has an optimal training result of 0.96 at epoch 250. For
Ponds 1 and 2, the best Adam test results are 0.9936 and 0.9975, respectively. Except
for AdaGrad, which yielded the lowest scores, the results show that Adam-based
optimizers gave the best results. Depending on the surroundings, an aquaponics system
with a well-optimized FNN may help make more accurate predictions of fish
development.
Keywords: Aquaponics, Feedforward neural network, Optimizer variants, Predictive analytics.
