Modelling Autonomous Agents’ Decisions in Learning to Cross a Cellular Automaton-Based Highway via Artificial Neural Networks

Abstract

<jats:p>A lot of effort has been devoted to mathematical modelling and simulation of complex systems for a better understanding of their dynamics and control. Modelling and analysis of computer simulations outcomes are also important aspects of studying the behaviour of complex systems. It often involves the use of both traditional and modern statistical approaches, including multiple linear regression, generalized linear model and non-linear regression models such as artificial neural networks. In this work, we first conduct a simulation study of the agents’ decisions learning to cross a cellular automaton based highway and then, we model the simulation data using artificial neural networks. Our research shows that artificial neural networks are capable of capturing the functional relationships between input and output variables of our simulation experiments, and they outperform the classical modelling approaches. The variable importance measure techniques can consistently identify the most dominant factors that affect the response variables, which help us to better understand how the decision-making by the autonomous agents is affected by the input factors. The significance of this work is in extending the investigations of complex systems from mathematical modelling and computer simulations to the analysis and modelling of the data obtained from the simulations using advanced statistical models.</jats:p>

Utilizing a multilayer perceptron artificial neural network to assess a virtual reality surgical procedure

“…Early stopping was frequently used in training the models, the optimal model stopped training after 3000 epochs as the validation loss started to slightly increase (Figure 3). In this work, the Algorithm was adapted to a multilayer neural network to calculate the Connection Weights Product (CWP) as recently suggested by multiple studies [24,25]. More specifically, this study adapted the algorithm to a two hidden layer network as demonstrated by Where 𝐶𝑊𝑃 𝑥,𝑧 is the connection weight product of an input metric to a class output , 𝑤 𝑥𝑛 is the weight connecting an input metric to a first hidden layer neuron 𝑛, 𝑣 𝑛𝑚 is the weight connecting a first hidden layer neuron 𝑛 to a second hidden layer neuron 𝑚, and 𝑞 𝑚𝑧 is the weight connecting a second hidden neuron 𝑚 to an output .…”