Simon Enjalbert

A Benefit/Cost/Deficit (BCD) model for learning from human errors

This paper proposes an original model for interpreting human errors, mainly violations, in terms of benefits, costs and potential deficits. This BCD model is then used as an input framework to learn from human errors, and two systems based on this model are developed: a case-based reasoning system and an artificial neural network system. These systems are used to predict a specific human car driving violation: not respecting the priority-to-the-right rule, which is a decision to remove a barrier. Both prediction systems learn from previous violation occurrences, using the BCD model and four criteria: safety, for identifying the deficit or the danger; and opportunity for action, driver comfort, and time spent; for identifying the benefits or the costs. The application of learning systems to predict car driving violations gives a rate over 80% of correct prediction after 10 iterations. These results are validated for the non-respect of priority-to-the-right rule.

How to learn from the resilience of Human-Machine Systems?

This paper proposes a functional architecture to learn from resilience. First, it defines the concept of resilience applied to Human-Machine System (HMS) in terms of safety management for perturbations and proposes some indicators to assess this resilience. Local and global indicators for evaluating human-machine resilience are used for several criteria. A multi-criteria resilience approach is then developed in order to monitor the evolution of local and global resilience. The resilience indicators are the possible inputs of a learning system that is capable of producing several outputs, such as predictions of the possible evolutions of the systems resilience and possible alternatives for human operators to control resilience. Our system has a feedback-feedforward architecture and is capable of learning from the resilience indicators. A practical example is explained in detail to illustrate the feasibility of such prediction.

Assessment of Transportation System Resilience

A transportation system like tramway or train is a system in which the functions of the human and the machine are interrelated and necessary for the operation of the whole system according to Human–Machine System (HMS) definition. Both human and machines are sources of system reliability and causes of accident occurrences. Considering the human behaviour contribution to HMS resilience, resilience can only be diagnosed if the human actions improve the system performances and help to recover from instability. Therefore, system resilience is the ability for a HMS to ensure performances and system stability whatever the context, i.e. after the occurrence of regular, unexpected or unprecedented disturbances. The COR&GEST platform is a railway simulation platform developed in the LAMIH in Valenciennes which involves a miniature railway structure. In order to study the human behaviour during the train driving activities with or without any technical failure occurrences, an experimental protocol was built with several inexperienced human operators. In railway transportation systems, traffic safety is the main performance criterion to take into account. Based on this criterion, authors propose to evaluate an instantaneous resilience indicator in order to assess the “local resilience” of HMS. As others performance criteria must be aggregated to reflect the whole studied HMS performance, the “global resilience” of HMS will be defined.

How to learn from the resilience of human-machine systems?

Abstract In this paper, we aim to analyse the resilience of Human-Machine Systems (HMS) in order to improve it from learning process. A State of Art is achieved and resilience engineering of HMS is defined. Then, human-machines’ learning processes supposed to improve systems’ resilience and indicators proposed in the literature to assess it are analysed. A perspective can be to propose an efficient indicator, for instance based on Benefit-Cost-Deficit (BCD) model, which can lead to the system resilience characterisation.

A hybrid reinforced learning system to estimate resilience indicators

Abstract This paper describes a learning system based on resilience indicators. It proposes a hybrid learning system to estimate Human–Machine System performance when facing unprecedented situations. Collected data from various criteria are compared with data estimated using the local and the global resilience indicators, to give both instantaneous and over-time Human–Machine System states. The learning system can be composed of two different, separate reinforcement functions; the first allowing reinforcement of its own system knowledge and the second allowing reinforcement of its estimation function. When used together in a hybrid approach, the resilience indicator estimation should be improved. The learning system is then applied in a simulated air transport context and the impact of each reinforcement function on resilience indicator estimation is assessed. The hypothesis on performance of hybrid reinforcement learning is confirmed and it provides better results than those obtained by the knowledge based reinforcement or the estimation based reinforcement alone.

A State of the Art in Feedforward-Feedback Learning Control Systems for Human Errors Prediction

Abstract In this paper, authors propose an overview of feedforward-feedback learning control systems that can be adapted for human errors prediction. A State of the Art in existing approaches for machines of feedback and/or feedforward learning control systems is presented and a synthesis relevant for prediction purposes is detailed. The possible application for learning systems based on human errors applied to Human Machine System (HMS) is then identified. A feedforward-feedback learning system applied to car driving simulation in order to predict intentional human errors is proposed. The paper concludes on relevant perspectives for feedforward-feedback learning systems to predict human errors and to increase HMS resilience facing unplanned disruptions in transportation.

Validation of a Unified Model of Driver behaviour for train domain

Abstract In this paper, authors goal is to present part of the work carried out within the European project ITERATE. The objective of ITERATE is to develop a Unified Model of Driver (UMD) behaviour and driver interaction with innovative technologies in emergency situations. Such a UMD could be used when designing innovative technologies for assessment and tuning of the systems in a safe and controllable environment, when guiding designers in identifying potential problem areas, when adapting systems to the drivers before being available on the market and providing better support to the driver in emergency situations, or, for authorities, as a guide in assessing and approving innovative technologies. Authors will mainly focus on the UMD validation process on a miniature railway platform. An exhaustive presentation of the COR&GEST platform is given to understand how the comparison with a numerical simulation is possible. The command of the miniature train model is demonstrated and the visual interface for driver is depicted. Then, the data collected during experiments and the process to analyse the results are indicated. Finally, the numerical SiMUD tool and the experimental results are compared to assess the validity of the UMD.

Toward an on-line and non-obtrusive workload assessment method

Abstract Workload is an index introduced in the 70th for ergonomic purposes, for evaluating the adequacy of tasks to the human operator abilities. Methods based on a self evaluation such as SWAT and TLX gave the best results, but mainly for assessing a total workload after the task has been performed. But in highly dynamical systems as transportation, the driver (or pilot) abilities can be enhanced by driving assistance tools on-line. Therefore new challenges appear which needs new methods for assessing Workload on-line, in real time and without disturbing the human operator. At LAMIH, an on-***line assessment method has been developed and partially validated by Millot in the past. This method is based on one hand, on two workload generators: temporal demands (time pressure) and functional demands (task difficulties). On the other hand, to cope with a possibility of assessment on-line, the workload is defined through the analogy with the physical notions of “power” and “energy”. The instantaneous workload Wl(t) is seen as the “power” the human operator invests on-line in the task in order to cope with the task demands. After a time available denoted TA, the human operator has spent a quantity of “energy” WL defined as the sum of the successive instantaneous Wl(t) along TA. These ideas have been validated in multitask situations for discrete as well as continuous tasks like driving tasks, but especially with temporal demands. This paper first compares several methods with the LAMIHs method. It then proposes an extension of this LAMIHs method in order to cope with the new dynamical constraints. Finally it proposes experimental protocols for validating the new LAMIHs method by comparing it with SWAT and TLX methods.

Observer-Based Tracking design using H∞ criteria: Application to eco-driving in a tramway system

Abstract This paper investigates an H∞ Observer-Based Controller design for tracking a tramway system eco-driving trajectory. The model of the tramway system is given in state space form, and the poor manoeuvres of the of the driver when following a reference trajectory are modeled as disturbances with finite energy that affect the system dynamics. To minimize the impact of poor driver manoeuvres, an H∞ Observer-Based Tracking Controller (H∞-OBTC) was designed and its conditions of existence are given. In addition, to ensure the robust convergence of the estimation and the tracking errors simultaneously, a new sufficient condition was obtained based on the Bounded Real Lemma. Two algorithms are presented to solve the robust stability condition obtained. The first one is based on a two-step procedure. Then a linearization approach was used to present the robust stability condition of the errors as a convex optimization problem with a Linear Matrix Inequality (LMI) constraint. The gain matrices of the H∞-OBTC can be computed by solving the LMI given, subject to a minimization constraint.