Applying Neural-Symbolic Cognitive Agents in Intelligent Transport Systems to reduce CO 2 emissions Leo de Penning, Artur S. d’Avila Garcez, Luis C. Lamb, Arjan Stuiver, and John-Jules Ch. Meyer Abstract—Providing personalized feedback in Intelligent Transport Systems is a powerful tool for instigating a change in driving behaviour and the reduction of CO 2 emissions. This requires a system that is capable of detecting driver character- istics from real-time vehicle data. In this paper, we apply the architecture and theory of a Neural-Symbolic Cognitive Agent (NSCA) to effectively learn and reason about observed driving behaviour and related driver characteristics. The NSCA archi- tecture combines neural learning and reasoning with symbolic temporal knowledge representation and is capable of encoding background knowledge, learning new hypotheses from observed data, and inferring new beliefs based on these hypotheses. Furthermore, it deals with uncertainty and errors in the data using a Bayesian inference model, and it scales well to hundreds of thousands of data samples as in the application reported in this paper. We have applied the NSCA in an Intelligent Transport System to reduce CO 2 emissions as part of an European Union project, called EcoDriver. Results reported in this paper show that the NSCA outperforms the state-of-the-art in this application area, and is applicable to very large data. Index Terms—Neural-Symbolic Learning and Reasoning, Driver modelling, Deep Learning, Restricted Boltzmann Ma- chines (RBM), I. I NTRODUCTION R ESEARCH has shown that providing feedback is a powerful tool for instigating a behaviour change (e.g. [1], [2], [3]). As part of an EU project on the reduction of CO 2 emissions, called EcoDriver, we focused on innovations in feedback advice strategies and Human-Machine Interface (HMI) solutions to maximize system effectiveness and accept- ance of an Intelligent Transport System that supports drivers to reduce their CO 2 emissions. For example, we applied an adaptive feedback strategy to tailor the feedback and HMI, based on the driving style of the driver (see Figure 1). This requires an automated driver type detection system that is able to reason online about observed driving behaviour and classify this behaviour in terms of several characteristics of the driver (e.g. learning vs. performance oriented, social vs. self oriented) using large amounts of real-time and noisy vehicle Leo de Penning is with the Department of Earth, Life and Social Sciences, TNO, Soesterberg, The Netherlands, e-mail: leo.depenning@tno.nl. Artur S. d’Avila Garcez is with the Department of Computer Science, City University London, UK, e-mail: a.garcez@city.ac.uk. Luis C. Lamb is with the Instituto de Informatica, UFRGS, Porto Alegre, RS, Brazil, e-mail: LuisLamb@acm.org. Arjan Stuiver is with the Department of Earth, Life and Social Sciences, TNO, Soesterberg, The Netherlands, e-mail: arjan.stuiver@tno.nl John-Jules Ch. Meyer is with the Department of Information and Computing Sciences, Universiteit Utrecht, Netherlands, e-mail: J.J.C.Meyer@uu.nl. data (e.g. steering angle, speed, rpm, gear, brake). Although the use of semi-controlled environments, like an instrumented car, simplifies the data and knowledge acquisition, the building of a model or intelligent autonomous agent that is capable of dealing with the many complex temporal relations in the observed data is a very difficult task, in particular in dynamic and non-stationary environments [4], [5]. Figure 1. EcoDriver human-machine interface for personalized feedback on CO 2 emissions. In this paper, we address the problems by applying the architecture and theory of the Neural-Symbolic Cognitive Agent (NSCA) described in [6], [7] to develop a robust model for driver type detection from real-time vehicle data. We will show the NSCA is capable of deep learning and reasoning about complex dynamic temporal relations, but also represent an agent’s knowledge in symbolic form for explanation and validation purposes, describing its decisions and providing feedback to the user. This is achieved by taking advantage of neural-symbolic integration [8], using the networks for performing robust learning and adaptation, and symbolic knowledge extraction for representing the temporal relations explicitly and for qualitative reasoning. We also provide additional proof showing the NSCA is capable of modelling any temporal logic program. The result is an agent model that is capable of efficient online learning and reasoning in complex, dynamic and non-stationary environments. We discuss the effective use and results of the NSCA as part of the driver type detection system, and refer to several other real- world applications of the NSCA (e.g. automated assessment