When training a feedforward stochastic gradient descendent trained neural network, there is a possibility of not learning a batch of patterns correctly that causes the network to fail in the predictions in the areas adjacent to those patterns. This problem has usually been resolved by directly adding more complexity to the network, normally by increasing the number of learning layers, which means it will be heavier to run on the workstation. In this paper, the properties and the effect of the patterns on the network are analysed and two main reasons why the patterns are not learned correctly are distinguished: the disappearance of the Jacobian gradient on the processing layers of the network and the opposite direction of the gradient of those patterns. A simplified experiment has been carried out on a simple neural network and the errors appearing during and after training have been monitored. Taking into account the data obtained, the initial hypothesis of causes seems to be correct. Finally, some corrections to the network are proposed with the aim of solving those training issues and to be able to offer a sufficiently correct prediction, in order to increase the complexity of the network as little as possible.

There is presently a need for more robust navigation algorithms for autonomous industrial vehicles. These have reasonably guaranteed the adequate reliability of the navigation. In the current work, the stability of a modified algorithm for collision-free guiding of this type of vehicle is ensured. A lateral control and a longitudinal control are implemented. To demonstrate their viability, a stability analysis employing the Lyapunov method is carried out. In addition, this mathematical analysis enables the constants of the designed algorithm to be determined. In conjunction with the navigation algorithm, the present work satisfactorily solves the localization problem, also known as simultaneous localization and mapping (SLAM). Simultaneously, a convolutional neural network is managed, which is used to calculate the trajectory to be followed by the AGV, by implementing the artificial vision. The use of neural networks for image processing is considered to constitute the most robust and flexible method for realising a navigation algorithm. In this way, the autonomous vehicle is provided with considerable autonomy. It can be regarded that the designed algorithm is adequate, being able to trace any type of path.

The Deep Deterministic Policy Gradient (DDPG) algorithm is a reinforcement learning algorithm that combines Q-learning with a policy. Nevertheless, this algorithm generates failures that are not well understood. Rather than looking for those errors, this study presents a way to evaluate the suitability of the results obtained. Using the purpose of autonomous vehicle navigation, the DDPG algorithm is applied, obtaining an agent capable of generating trajectories. This agent is evaluated in terms of stability through the Lyapunov function, verifying if the proposed navigation objectives are achieved. The reward function of the DDPG is used because it is unknown if the neural networks of the actor and the critic are correctly trained. Two agents are obtained, and a comparison is performed between them in terms of stability, demonstrating that the Lyapunov function can be used as an evaluation method for agents obtained by the DDPG algorithm. Verifying the stability at a fixed future horizon, it is possible to determine whether the obtained agent is valid and can be used as a vehicle controller, so a task-satisfaction assessment can be performed. Furthermore, the proposed analysis is an indication of which parts of the navigation area are insufficient in training terms.

The stochastic gradient descendant algorithm is one of the most popular neural network training algorithms. Many authors have contributed to modifying or adapting its shape and parametrizations in order to improve its performance. In this paper, the authors propose two modifications on this algorithm that can result in a better performance without increasing significantly the computational and time resources needed. The first one is a dynamic learning ratio depending on the network layer where it is applied, and the second one is a dynamic drop-out that decreases through the epochs of training. These techniques have been tested against different benchmark function to see their effect on the learning process. The obtained results show that the application of these techniques improves the performance of the learning of the neural network, especially when they are used together.

Regarding autonomous vehicle navigation, reinforcement learning is a technique that has demonstrated significant results. Nevertheless, it is a technique with a high number of parameters that need to be optimised without prior information, and correctly performing this is a complicated task. In this research study, a system based on the principles of reinforcement learning, specifically on the concept of rewards, is presented. A mathematical expression was proposed to control the vehicle’s direction based on its position, the obstacles in the environment and the destination. In this equation proposal, there was only one unknown parameter that regulated the degree of the action to be taken, and this was optimised through the genetic algorithm. In this way, a less computationally expensive navigation algorithm was presented, as it avoided the use of neural networks. The controller’s time to obtain the navigation instructions was around 6.201·10⁻⁴ s. This algorithm is an efficient and accurate system which manages not to collide with obstacles and to reach the destination from any position. Moreover, in most cases, it has been found that the proposed navigations are also optimal.