Using Neural Networks to Improve Behavioural Realism in Driving Simulation Scenarios

This paper describes the development of a neural network driver agent to improve the realism and perceived “intelligence” of autonomous vehicles in driving simulation scenarios. Driver agent refers to simulated entities which have an internal representation or knowledge of the traffic environment through their vision and can determine their own actions using their decision/cognitive capabilities. Driver agents can therefore mimic the basic elements of human drivers, for example planning, perception, learning and decision making, and therefore display behavioural intelligence. Neural networks are a modelling technique that can be used to improve the behavioural intelligence of driver agents. Initial data was collected from human drivers in a TRL passenger car simulator to develop and train a feed-forward multi-layer neural network and enable the neural driver agent to learn how to produce lane changing behaviour. The behaviour of the neural driver agent and human drivers were then compared and it was found that the neural driver agent produced good results when estimating the change in direction and speed required for lane changing manoeuvres. The behavioural realism of the neural driver agent was assessed and it was found that participants could not correctly distinguish between human and neural driver agent driven vehicles, suggesting that the neural driver agent is capable of improving participants’ immersion in simulator scenarios.