- Email: [email protected]
AUTONOMY CONTROL OF HUMAN-MACHINE SYSTEMS
AUTONOMY CONTROL OF HUMAN-MACHINE SYSTEMS F. Vanderhaegen1,2,3 1
Univ Lille Nord de France, F-59000 Lille, France UVHC, LAMIH, F-59313 Valenciennes, France 3 CNRS, FRE 3304, F-59313 Valenciennes, France Email : [email protected] 2
Abstract: this paper proposes a prospective discussion on the concepts of decisional autonomy and its control such as “autonomation” or “autonomisation”, applied to humanmachine systems. The autonomy of a system requires decisional capacities: capacities to identify a lack of knowledge that requires the definition of an allocation of activities between decision makers, capacities to achieve the allocated activities, capacities to learn from known and unknown situations. They are interpreted in terms of know-how, knowhow-to-cooperate and know-how-to-learn. Cooperation between human and machine involves usually static knowledge and is also a way for a decision maker to learn from the behaviours of the other ones. The paper proposes several ways to manage or refine the system knowledge dynamically, and discusses on possible learning capacities. It details a model based on the diagnosis, the prognosis and the trial-and-error functions to optimize the control of known situations or to create new parades facing new situations. A feedforward-feedback system based model is proposed for implementing the autonomy control concepts to automated support tools of human-machine systems. Copyright © 2010 IFAC Keywords: autonomy control, feedforward-feedback learning, cooperation, human factors, human error, learning systems. 1. INTRODUCTION The limited cognitive capacities of automated systems lead their designers to develop solutions in order to make them more autonomous. These future systems will then have capacities of learning to manage their knowledge dynamically and to increase their decisional capacities. The on-line management of the limited capacity of decision makers involved in the control of a given human-machine system can be done by making them cooperate. Several learning strategies can then be applied to make people or machines more autonomous. Individual strategies are done via observations of internal and external phenomena. They are strategies such as the learning from human errors, the learning by imitation by copying observed behaviours or phenomena, the learning by trial-error to manage the uncertainty, etc. Collective strategies for learning relates to interaction between decision makers. They are cooperation based strategies in order to facilitate intentionally the activity of the decision makers and to make them more autonomous. The last point requires pedagogical cooperation. This paper is a contribution for the automation of the autonomy of automated support tools implemented into human-machine systems. It treats several concepts related to system autonomy: mode of
autonomy, cooperation, and learning. It proposes a so-called know-how-to-learn that is the capacity of decision makers to learn from the control of known and unknown situations. Finally, it details a feedforward-feedback learning system to be implemented into automated support systems. 2. AUTONOMY AND MODES OF AUTONOMY The autonomy concept is used in different domains such as manufacturing, medicine, psychology, sociology, education or engineering. Communities such as medicine, sociology or psychology focus on the behavioural autonomy or on the so-called “autonomisation” of people. It is then the capacity of dependent persons to become independent (Lemétayer, Lanfranchi, 2006). This independency can relate to different criteria: physical criterion, financial criteria, decisional criterion, etc. With regard to educational viewpoint, the behavioural autonomy control process aims at reducing the external assistance for learning (Carton, 1984). A minimum degree of liberty is then allocated to human in order to learn from the process they control as they want (Carton, 1984). Other communities such as engineering or manufacturing apply the so-called concept of autonomation, the English translation of the Japanese word “jidoka”. Autonomation concerns the transfer of human cognition to machine proposing a set of
human or technical systems to detect non-conformity events and to stop a process in order to limit the creation or the propagation of wrong products (Saurin et al., 2008). It relates to the autonomous control of system degradations for maintaining quantity and quality of manufactured products (Xiaobo, Ohno, 2000; Black, 2002). To make a human or a machine autonomous, several characteristics are required (Demazeau, 1995; Carabelea and Boissier, 2006; Zieba et al., 2009): • An agent has several available behaviours and can choose one of them to achieve a goal. • An agent evolves in a given environment but it is not controlled by this environment. • An agent manages the interactions with other agents and can accept or refuse a goal proposed by another agent. • An agent has the capacity to violate coordination process rules. • An agent can choose an action without interacting with other agents. • An agent requires knowledge, physical and cognitive capacities to apply its knowledge, and predefined prescriptions to authorize or prohibit the achievement of some actions. • With these three levels (i.e., knowledge, physical and cognitive capacities and prescriptions), an agent is capable to achieve a given goal, to select an alternative and to define an action plan. Depending on the limited capacities of a human or a machine related to these characteristics, modes of autonomy can be implemented. They are linked with the degree of automation and the task allocation between human and machine implemented into a human-machine system (Sheridan, 1999; Inagaki et al., 1998; Zieba et al., 2009). Human-machine cooperation concept can then facilitate the interactions between human and machine in order to recover a lack of autonomy of a decision-maker. 3. AUTONOMY AND COOPERATION The decisional autonomy relates the Know-how of the decision makers involved in the control of a given human-machine system. The limited capacity of these decision makers leads them to cooperate Definition of cooperative agents: implication for autonomy In the field of cognitive psychology, Hoc (1996) and Millot & Hoc (1997) have proposed the following definition: "two agents are cooperating if 1) each one strives towards goals and can interfere with the other, and 2) each agent tries to detect and process such interference to make the other's activities easier".
Cooperation between agents implies a bi-directional interaction and agents or decision makers can also cooperate when each one aims at facilitating their own activities (Vanderhaegen, 1999, 2003). As a matter of fact, the agents of a given group cooperate together in order to facilitate the activities of the agents of other groups or/and the activity of their own group. The human-machine cooperation can then be applied for both individual and collective interests. The decisional autonomy can then be managed by applying the human-machine cooperation structures. This requires both a communication and allocation systems (Vanderhaegen, 1997, 1999a). The allocation system shares the resources (e.g. the machines, humans, or interfaces), the activities of these resources (e.g., functions, tasks, or actions) or the support for their activities (e.g., information, data, knowledge) in order to recover a lack of autonomy between decision makers. The communication system aims at making this sharing possible and optimal. For instance, it can lead to minimize possible errors of the decision makers, to optimize their mutual understanding and their joint decision or to confirm a potential alternative in case of doubt. Several strategies can be applied for managing the human decisional autonomy. Even if human operators have sufficient knowledge to control a given process, their autonomy can depend on their workload evolution (Vanderhaegen, 1999b, 1999c). The risk of human error increases when the workload is low (i.e. underload that may generate situations of hypovigilance) or high (i.e. overload that may generate situations of hyper-vigilance or of panic). The control of the autonomy of a human-machine system depends on several constraints. These constraints relate to the characteristic for the possible sharing of the resources, of the activities of these resources or of the supports for these activities. For instance, regarding allocation of a function, of a task or of an action, they are constraints such as: • The possible decomposition of a function, a task or an action into sub-processes (i.e. sub-functions, sub-tasks or sub-actions). • The possible allocation of a function, a task, an action or their associated processes to human and machine. • The interruption of the achievement of a function, a task, an action or their associated sub-processes. • The possible recovery of an initial allocation of a function, a task, an action or their associated subprocesses. • The possible recovery of an erroneous function, task, action or associated sub-processes. • Etc. Different allocation control modes are available for controlling such a decisional autonomy of a humanmachine system (Vanderhaegen, 2003). The development of cooperative automated system requires the implementation of two classes of
capacities (Millot, 1998; Millot and Vanderhaegen, 2008): • The know-how that relates to the knowledge of the system. • The know-how-to-cooperate that makes cooperation activities possible. Nevertheless, these structures of human-machine cooperation applied for the decisional autonomy control of human-machine system do not integrate the possible evolution of the system knowledge, i.e. they focus on the prescription of a static know-how of the automated support tools without considering any dynamic process for managing initial know-how to optimize system behaviour or for creating new knowhow to control new situations. Moreover, these approaches do not consider the possible evolution of the human behaviour when they use these automated support systems. For instance, the use of a technical system by human operators can transform initial cooperative activities into competitive ones (Vanderhaegen et al., 2006). A so-called know-how-to-learn based capacity should be then introduced to manage dynamically the evolution of this know-how (including the know-howto-cooperate), and to increase the decisional autonomy control of human-machine systems. 4. AUTONOMY AND LEARNING The control of the autonomy of a human-machine system requires the application of two main strategies: • The autonomy based on the management of static knowledge. It consists in controlling the optimal balancing between the decisional autonomy of human operators and automated systems in order to make the global human-machine system autonomous. This implies the development of the concepts of the human-machine cooperation proposed above and the use of the static knowledge implemented into the automated support tools. • The autonomy based on the management of dynamic knowledge. It consists in controlling dynamically the autonomy of a given system by the system itself or by other systems. Two concepts may be applied: o The autonomation, i.e., the automation of the autonomy to manage degraded or new situations and to transfer the human operator’s knowledge into the machine’s one, or viceversa. o The autonomisation, i.e. the capacities of a given system to make this autonomation possible. For instance, these capacities can relate to the auto-learning based capacities or to the learning capacities based on cooperation between decision makers or to the learning capacities based on error occurrence and management processes.
The autonomy of a human-machine system composed by several decision makers includes different cognitive capacities such as: • Capacities to define and to activate the allocation of activities between decision makers. These capacities aim at managing the lacks of the knowhow of the decision makers. Some of them relate to the know-how-to-cooperate and their management can require different hierarchical decisional levels and evolve from strategic allocation decisions to tactical or real-time allocation ones (Vanderhaegen, 2003). • Capacities to achieve these allocated activities. These capacities relate to the know-how of the decision makers. • Capacities to refine knowledge and to manage new knowledge. These capacities relate to the know-how-to-learn and they allow the dynamic management of knowledge. The human modelling of such a know-how-to-learn involves the capacities to identify the state of a given process in order to produce hypotheses on its evolution, by applying the diagnosis, the prognosis and/or the trial-and-error functions, Figure 1. The prognosis function leads to the identification of the possible evolutions of the system state with or without actions. The diagnosis function relates to the explanation of the current system state regarding the previous ones. When this identification is not possible because of the occurrence of unknown system states, a so-called trial-and-error process is required in order to apply an action on the process and to wait and see its consequences. This trial-anderror process aims at trying to understand the current system state and to propose a new adapted action plan. The prognosis, the diagnosis and the trial-anderror based processes can be managed or optimized individually or collectively, i.e. they can be shared between decision makers or be managed together in order to recover a lack of know-how. This leads to the individual or collective dynamic management of knowledge and to the improvement of the learning capacities. New action plan definition with planned consequences
Application of a predefined action plan
Control and supervision of the impact of an action to the system
Selection of the more appropriate alternative
Hypothesis on the system state and its evolution Yes
Yes
“Wait and see” process
Occurrence of an unforeseen state
Identification of system state possible?
Prognosis related to the future states possible? Yes
No
Application of an action without planned consequences
No
Diagnosis related to the previous states possible?
No
Fig. 1. The diagnosis, prognosis and trial-and-error processes.
The technical implementation of such a modelling of the know-how-to-learn requires the development of learning mechanisms and of knowledge management (Vanderhaegen et al., 2009). Among these mechanisms, there is the feedforward-feedback based mechanism that consists in using the current knowledge related to previous activities in order to calculate the future ones, Figure 2. Feedback process on the N previous iterations
FBS: Feedback based system FFS: Feedforward based system HO: Human operator
ei-1
u i-1
Iteration i FBS
uˆ i
FFS
Feedforward process on the current and next M iterations
ei Iterations
i-N
i-2
i-1
i+1
i+M
HO
i (current iteration) Feedback process on the N previous iterations
ui
Process
Feedforward process on the current and next M iterations Iterations
i-N-1
i-1
i
i+2
i+M+1
i+1 (current iteration)
Iteration i+1 FBS
Fig. 2. The feedforward-feedback learning based process.
uˆ i+1
FFS
The feedforward process aims at assessing the future possible decisions regarding the current system states and the management of the previous ones. It applies the static knowledge. The feedback aims at recovering possible erroneous knowledge, at refining knowledge or at creating new knowledge. The implementation of the diagnosis and the prognosis functions integrates one or several iterations whereas the trial-and-error one takes into account a single iteration (i.e., N=1 and M=0). An iteration relates to the current state of a given process, i.e., its associated data (noted e) and decisions (note u). The possible development of the know-how-to-learn of automated support tools may include both the feedforward and feedback based systems, Figure 3. For instance, the production of possible decisions (noted û ) by the automated support tools may be generated by the FFS with regard to the data e, and the comparison between the real human decisions (noted u) and the decisions planned by the automated system (noted û ) allows the dynamic management and refinement of the knowledge implemented into the FBS that integrates the data and the decisions of the previous iterations. The implementation of such know-how-to-learn modelling will aim at improving the autonomy of automated supports tools by managing their knowledge themselves or by interacting with human operators facing known or unknown situations.
ei+1 HO
u i+1
Process
ei+1
u i+1 Fig. 3. Toward a feedforward-feedback learning based structure. Future studies have to be realized in order to assess the feasibility of such an implementation, and to compare different structures of the feedforwardfeedback based systems in order to select the more appropriate one for improving knowledge on known situations and for creating knowledge related to new situations.
5. CONCLUSION
REFERENCES
The increasing of the decisional autonomy of a human-machine system consists in using the capacities of all the decision makers in order to make the control of any operational situations possible. This control can be done by sharing human and machine resources, by sharing the activities related to these resources or by sharing the support for these activities.
Métayer, F. and Lafranchi, J.-B. (2006). Mental deficiency, behavioural autonomy and problemsolving in early period: factorial analysis of parent-child behaviours with non-defective and Down syndrome toddler. Revue Européenne de Psychologie Appliquée, 56, 179-189. Black, J. T. (2002). Design for system success. Journal of Manufacturing Systems, 20, 6, 1-6. Carabelea, C. and Boissier, O. (2006). Coordinating agents in organisations using social commitments. Electronic Notes in Theoretical Computer Science, 150, 73-91. Carton, F. (1984). Autonomy behaviour systems of language learning (Systèmes autonomisants d’apprentissage des langues). Paper presented at the workshop “Autonomy and Teaching French” at the 6th World Congress of the International Federation of Professors of French (FIPF), 15-20 July 1984, Québec, Canada. Demazeau, Y. (1995). From interactions to collective behaviour in agent-based systems. Proceedings of the 1st European Conference on Cognitive Science, Saint-Malo, France, pp. 117-132. Hoc, J.-M (1996). Supervision et Contrôle de Processus: la Cognition en Situation Dynamique. (Supervision and control of process : cognition in dynamic situation). Presses Universitaires de Grenoble, Grenoble, France. Inagaki, T., Moray, N., Itoh, M. (1998). Trust self confidence and authority in human-machine systems. Proceedings of the 7th IFAC/IFIP/IFORS/IEA Symposium on Analysis, Design, and Evaluation of Human-Machine Systems, September 16-18 1998, Kyoto, Japan, pp. 431-436. Millot, P. (1998). Concepts and limits for HumanMachine Cooperation. Proceedings of the IEEE SMC CESA'98 Conference, Hammamet, Tunisia, April. Millot, P. and Hoc, J.-M. (1997). Human-machine cooperation: metaphore or possible reality? Proceedings of the 2nd European Conference on Cognitive Science, April 9-11, Manchester, UK, pp.165-174. Millot, P. and Vanderhaegen, F. (2008). Toward cooperatve and human error-tolerant system. Proceedings of the 17th IFAC World Congress, 611 June 2008, Seoul, Korea. Polet, P., F. Vanderhaegen and R. Amalberti (2003). Modelling Border-line tolerated conditions of use (BTCUs) and associated risks. Safety Science, 41, 111-136. Polet, P., Vanderhaegen, F., Wieringa, P.(2002). Theory of safety related violation of system barriers. Cognition Technology & Work, 4, 3, 171-179 Saurin T. A., Formosao, C. T., Cambraia, F. B. (2008). An analysis of construction safety best practices from a cognitive engineering systems perspective. Safety Science, 46, 1169-1183.
However, the feasibility of applying such humanmachine cooperation principles is often limited to the study of static knowledge implemented on the automated decision support tools. Indeed, the cognitive capacity of automated system is usually limited and cannot evolve dynamically. This paper has then proposed an original view on the decisional autonomy control process applied for human-machine systems. It developed the so-called know-how-tolearn in order to manage dynamically the knowledge of the automated tools and to increase the system autonomy. The proposed approach focused toward a feedforward-feedback learning based process that implement the activities Future development will study the feasibility of managing dynamically the automated system autonomy in terms of cognitive capacities to make decisions possible facing know and unknown situations. Such future systems will be able to autolearn from human errors, from interaction with humans, etc. Cooperation principles will be then applied not only for adjusting the human-machine system autonomy by sharing physical resources, resource activities or activity supports between human and machine, but also for increasing the decisional autonomy of the implemented automated decision support tools. Automated tools with such learning capacities will be able to build a model of their users, to identify possible diverted uses and to control the associated risks of such wrong uses. ACKNOWLEDGEMENTS The present research work has been supported by the European Research Group on Human-Machine System in Transportation (GDR E HAMASYT), the International Campus on Safety and Intermodality in Transportation (CISIT), the European Community, the Délégation Régionale à la Recherche et à la Technologie, the Ministère de l'Enseignement Supérieur et de la Recherche, the Région Nord Pas de Calais and the Centre National de la Recherche Scientifique. The authors gratefully acknowledge the support of these institutions.
Sheridan, T. B. (2000). Function allocation algorithm, alchemy or apostasy. International Journal of Human-Machine Studies, 52, 203-216. Vanderhaegen, F. (2004). The Benefit-Cost-Deficit (BCD) model for human analysis and control. Proceedings of the 9th IFAC/IFORS/IEA symposium on Analysis, Design, and Evaluation of Human-Machine Systems, Atlanta, GA, USA, 7-9 September 2004. Vanderhaegen, F., 1997. Multilevel organization design: the case of the air traffic control. Control Engineering Practice, 5, 3, 391-399. Vanderhaegen, F. 1999a. Cooperative system organisation and task allocation: illustration of task allocation in air traffic control. Le Travail Humain, 63, 3, 197-222. Vanderhaegen, F. (1999b). Multilevel allocation modes - Allocator control policies to share tasks between human and computer. System Analysis Modelling Simulation, 35, 191-213. Vanderhaegen, F. (1999c). Toward a model of unreliability to study error prevention supports. Interacting With Computers, 11, 575-595. Vanderhaegen, F. (2003). Analyse et contrôle de l’erreur humaine (Analysis and control of human error). Hermès Science Publications – Lavoisier, Paris, France. Vanderhaegen, F., Chalmé, S., Anceaux, F., Millot, P. (2006). Principles of cooperation and competition - Application to car driver behavior analysis. Cognition Technology & Work, 8, 3, 183-192. Vanderhaegen, F., Zieba, S., Polet, P. (2009). A reinforced iterative formalism to learn from human errors and uncertainty. Engineering Applications and Artificial Intelligence, 22, 654659. Xiabo, Z. and Ohno, K. (2000). Properties of a sequencing problem for a mixed model assembly line with conveyor stoppage. European Journal of Operational Research, 124, 560-570. Zieba, S., Vanderhaegen, F., Polet, P., Debernard, S. (2009). Resilience of human-robot system using ajustable autonomy and human-robot collaborative control. International Journal on Adaptative and Innovative, 1, 13-29.
Univ Lille Nord de France, F-59000 Lille, France UVHC, LAMIH, F-59313 Valenciennes, France 3 CNRS, FRE 3304, F-59313 Valenciennes, France Email : [email protected] 2
Abstract: this paper proposes a prospective discussion on the concepts of decisional autonomy and its control such as “autonomation” or “autonomisation”, applied to humanmachine systems. The autonomy of a system requires decisional capacities: capacities to identify a lack of knowledge that requires the definition of an allocation of activities between decision makers, capacities to achieve the allocated activities, capacities to learn from known and unknown situations. They are interpreted in terms of know-how, knowhow-to-cooperate and know-how-to-learn. Cooperation between human and machine involves usually static knowledge and is also a way for a decision maker to learn from the behaviours of the other ones. The paper proposes several ways to manage or refine the system knowledge dynamically, and discusses on possible learning capacities. It details a model based on the diagnosis, the prognosis and the trial-and-error functions to optimize the control of known situations or to create new parades facing new situations. A feedforward-feedback system based model is proposed for implementing the autonomy control concepts to automated support tools of human-machine systems. Copyright © 2010 IFAC Keywords: autonomy control, feedforward-feedback learning, cooperation, human factors, human error, learning systems. 1. INTRODUCTION The limited cognitive capacities of automated systems lead their designers to develop solutions in order to make them more autonomous. These future systems will then have capacities of learning to manage their knowledge dynamically and to increase their decisional capacities. The on-line management of the limited capacity of decision makers involved in the control of a given human-machine system can be done by making them cooperate. Several learning strategies can then be applied to make people or machines more autonomous. Individual strategies are done via observations of internal and external phenomena. They are strategies such as the learning from human errors, the learning by imitation by copying observed behaviours or phenomena, the learning by trial-error to manage the uncertainty, etc. Collective strategies for learning relates to interaction between decision makers. They are cooperation based strategies in order to facilitate intentionally the activity of the decision makers and to make them more autonomous. The last point requires pedagogical cooperation. This paper is a contribution for the automation of the autonomy of automated support tools implemented into human-machine systems. It treats several concepts related to system autonomy: mode of
autonomy, cooperation, and learning. It proposes a so-called know-how-to-learn that is the capacity of decision makers to learn from the control of known and unknown situations. Finally, it details a feedforward-feedback learning system to be implemented into automated support systems. 2. AUTONOMY AND MODES OF AUTONOMY The autonomy concept is used in different domains such as manufacturing, medicine, psychology, sociology, education or engineering. Communities such as medicine, sociology or psychology focus on the behavioural autonomy or on the so-called “autonomisation” of people. It is then the capacity of dependent persons to become independent (Lemétayer, Lanfranchi, 2006). This independency can relate to different criteria: physical criterion, financial criteria, decisional criterion, etc. With regard to educational viewpoint, the behavioural autonomy control process aims at reducing the external assistance for learning (Carton, 1984). A minimum degree of liberty is then allocated to human in order to learn from the process they control as they want (Carton, 1984). Other communities such as engineering or manufacturing apply the so-called concept of autonomation, the English translation of the Japanese word “jidoka”. Autonomation concerns the transfer of human cognition to machine proposing a set of
human or technical systems to detect non-conformity events and to stop a process in order to limit the creation or the propagation of wrong products (Saurin et al., 2008). It relates to the autonomous control of system degradations for maintaining quantity and quality of manufactured products (Xiaobo, Ohno, 2000; Black, 2002). To make a human or a machine autonomous, several characteristics are required (Demazeau, 1995; Carabelea and Boissier, 2006; Zieba et al., 2009): • An agent has several available behaviours and can choose one of them to achieve a goal. • An agent evolves in a given environment but it is not controlled by this environment. • An agent manages the interactions with other agents and can accept or refuse a goal proposed by another agent. • An agent has the capacity to violate coordination process rules. • An agent can choose an action without interacting with other agents. • An agent requires knowledge, physical and cognitive capacities to apply its knowledge, and predefined prescriptions to authorize or prohibit the achievement of some actions. • With these three levels (i.e., knowledge, physical and cognitive capacities and prescriptions), an agent is capable to achieve a given goal, to select an alternative and to define an action plan. Depending on the limited capacities of a human or a machine related to these characteristics, modes of autonomy can be implemented. They are linked with the degree of automation and the task allocation between human and machine implemented into a human-machine system (Sheridan, 1999; Inagaki et al., 1998; Zieba et al., 2009). Human-machine cooperation concept can then facilitate the interactions between human and machine in order to recover a lack of autonomy of a decision-maker. 3. AUTONOMY AND COOPERATION The decisional autonomy relates the Know-how of the decision makers involved in the control of a given human-machine system. The limited capacity of these decision makers leads them to cooperate Definition of cooperative agents: implication for autonomy In the field of cognitive psychology, Hoc (1996) and Millot & Hoc (1997) have proposed the following definition: "two agents are cooperating if 1) each one strives towards goals and can interfere with the other, and 2) each agent tries to detect and process such interference to make the other's activities easier".
Cooperation between agents implies a bi-directional interaction and agents or decision makers can also cooperate when each one aims at facilitating their own activities (Vanderhaegen, 1999, 2003). As a matter of fact, the agents of a given group cooperate together in order to facilitate the activities of the agents of other groups or/and the activity of their own group. The human-machine cooperation can then be applied for both individual and collective interests. The decisional autonomy can then be managed by applying the human-machine cooperation structures. This requires both a communication and allocation systems (Vanderhaegen, 1997, 1999a). The allocation system shares the resources (e.g. the machines, humans, or interfaces), the activities of these resources (e.g., functions, tasks, or actions) or the support for their activities (e.g., information, data, knowledge) in order to recover a lack of autonomy between decision makers. The communication system aims at making this sharing possible and optimal. For instance, it can lead to minimize possible errors of the decision makers, to optimize their mutual understanding and their joint decision or to confirm a potential alternative in case of doubt. Several strategies can be applied for managing the human decisional autonomy. Even if human operators have sufficient knowledge to control a given process, their autonomy can depend on their workload evolution (Vanderhaegen, 1999b, 1999c). The risk of human error increases when the workload is low (i.e. underload that may generate situations of hypovigilance) or high (i.e. overload that may generate situations of hyper-vigilance or of panic). The control of the autonomy of a human-machine system depends on several constraints. These constraints relate to the characteristic for the possible sharing of the resources, of the activities of these resources or of the supports for these activities. For instance, regarding allocation of a function, of a task or of an action, they are constraints such as: • The possible decomposition of a function, a task or an action into sub-processes (i.e. sub-functions, sub-tasks or sub-actions). • The possible allocation of a function, a task, an action or their associated processes to human and machine. • The interruption of the achievement of a function, a task, an action or their associated sub-processes. • The possible recovery of an initial allocation of a function, a task, an action or their associated subprocesses. • The possible recovery of an erroneous function, task, action or associated sub-processes. • Etc. Different allocation control modes are available for controlling such a decisional autonomy of a humanmachine system (Vanderhaegen, 2003). The development of cooperative automated system requires the implementation of two classes of
capacities (Millot, 1998; Millot and Vanderhaegen, 2008): • The know-how that relates to the knowledge of the system. • The know-how-to-cooperate that makes cooperation activities possible. Nevertheless, these structures of human-machine cooperation applied for the decisional autonomy control of human-machine system do not integrate the possible evolution of the system knowledge, i.e. they focus on the prescription of a static know-how of the automated support tools without considering any dynamic process for managing initial know-how to optimize system behaviour or for creating new knowhow to control new situations. Moreover, these approaches do not consider the possible evolution of the human behaviour when they use these automated support systems. For instance, the use of a technical system by human operators can transform initial cooperative activities into competitive ones (Vanderhaegen et al., 2006). A so-called know-how-to-learn based capacity should be then introduced to manage dynamically the evolution of this know-how (including the know-howto-cooperate), and to increase the decisional autonomy control of human-machine systems. 4. AUTONOMY AND LEARNING The control of the autonomy of a human-machine system requires the application of two main strategies: • The autonomy based on the management of static knowledge. It consists in controlling the optimal balancing between the decisional autonomy of human operators and automated systems in order to make the global human-machine system autonomous. This implies the development of the concepts of the human-machine cooperation proposed above and the use of the static knowledge implemented into the automated support tools. • The autonomy based on the management of dynamic knowledge. It consists in controlling dynamically the autonomy of a given system by the system itself or by other systems. Two concepts may be applied: o The autonomation, i.e., the automation of the autonomy to manage degraded or new situations and to transfer the human operator’s knowledge into the machine’s one, or viceversa. o The autonomisation, i.e. the capacities of a given system to make this autonomation possible. For instance, these capacities can relate to the auto-learning based capacities or to the learning capacities based on cooperation between decision makers or to the learning capacities based on error occurrence and management processes.
The autonomy of a human-machine system composed by several decision makers includes different cognitive capacities such as: • Capacities to define and to activate the allocation of activities between decision makers. These capacities aim at managing the lacks of the knowhow of the decision makers. Some of them relate to the know-how-to-cooperate and their management can require different hierarchical decisional levels and evolve from strategic allocation decisions to tactical or real-time allocation ones (Vanderhaegen, 2003). • Capacities to achieve these allocated activities. These capacities relate to the know-how of the decision makers. • Capacities to refine knowledge and to manage new knowledge. These capacities relate to the know-how-to-learn and they allow the dynamic management of knowledge. The human modelling of such a know-how-to-learn involves the capacities to identify the state of a given process in order to produce hypotheses on its evolution, by applying the diagnosis, the prognosis and/or the trial-and-error functions, Figure 1. The prognosis function leads to the identification of the possible evolutions of the system state with or without actions. The diagnosis function relates to the explanation of the current system state regarding the previous ones. When this identification is not possible because of the occurrence of unknown system states, a so-called trial-and-error process is required in order to apply an action on the process and to wait and see its consequences. This trial-anderror process aims at trying to understand the current system state and to propose a new adapted action plan. The prognosis, the diagnosis and the trial-anderror based processes can be managed or optimized individually or collectively, i.e. they can be shared between decision makers or be managed together in order to recover a lack of know-how. This leads to the individual or collective dynamic management of knowledge and to the improvement of the learning capacities. New action plan definition with planned consequences
Application of a predefined action plan
Control and supervision of the impact of an action to the system
Selection of the more appropriate alternative
Hypothesis on the system state and its evolution Yes
Yes
“Wait and see” process
Occurrence of an unforeseen state
Identification of system state possible?
Prognosis related to the future states possible? Yes
No
Application of an action without planned consequences
No
Diagnosis related to the previous states possible?
No
Fig. 1. The diagnosis, prognosis and trial-and-error processes.
The technical implementation of such a modelling of the know-how-to-learn requires the development of learning mechanisms and of knowledge management (Vanderhaegen et al., 2009). Among these mechanisms, there is the feedforward-feedback based mechanism that consists in using the current knowledge related to previous activities in order to calculate the future ones, Figure 2. Feedback process on the N previous iterations
FBS: Feedback based system FFS: Feedforward based system HO: Human operator
ei-1
u i-1
Iteration i FBS
uˆ i
FFS
Feedforward process on the current and next M iterations
ei Iterations
i-N
i-2
i-1
i+1
i+M
HO
i (current iteration) Feedback process on the N previous iterations
ui
Process
Feedforward process on the current and next M iterations Iterations
i-N-1
i-1
i
i+2
i+M+1
i+1 (current iteration)
Iteration i+1 FBS
Fig. 2. The feedforward-feedback learning based process.
uˆ i+1
FFS
The feedforward process aims at assessing the future possible decisions regarding the current system states and the management of the previous ones. It applies the static knowledge. The feedback aims at recovering possible erroneous knowledge, at refining knowledge or at creating new knowledge. The implementation of the diagnosis and the prognosis functions integrates one or several iterations whereas the trial-and-error one takes into account a single iteration (i.e., N=1 and M=0). An iteration relates to the current state of a given process, i.e., its associated data (noted e) and decisions (note u). The possible development of the know-how-to-learn of automated support tools may include both the feedforward and feedback based systems, Figure 3. For instance, the production of possible decisions (noted û ) by the automated support tools may be generated by the FFS with regard to the data e, and the comparison between the real human decisions (noted u) and the decisions planned by the automated system (noted û ) allows the dynamic management and refinement of the knowledge implemented into the FBS that integrates the data and the decisions of the previous iterations. The implementation of such know-how-to-learn modelling will aim at improving the autonomy of automated supports tools by managing their knowledge themselves or by interacting with human operators facing known or unknown situations.
ei+1 HO
u i+1
Process
ei+1
u i+1 Fig. 3. Toward a feedforward-feedback learning based structure. Future studies have to be realized in order to assess the feasibility of such an implementation, and to compare different structures of the feedforwardfeedback based systems in order to select the more appropriate one for improving knowledge on known situations and for creating knowledge related to new situations.
5. CONCLUSION
REFERENCES
The increasing of the decisional autonomy of a human-machine system consists in using the capacities of all the decision makers in order to make the control of any operational situations possible. This control can be done by sharing human and machine resources, by sharing the activities related to these resources or by sharing the support for these activities.
Métayer, F. and Lafranchi, J.-B. (2006). Mental deficiency, behavioural autonomy and problemsolving in early period: factorial analysis of parent-child behaviours with non-defective and Down syndrome toddler. Revue Européenne de Psychologie Appliquée, 56, 179-189. Black, J. T. (2002). Design for system success. Journal of Manufacturing Systems, 20, 6, 1-6. Carabelea, C. and Boissier, O. (2006). Coordinating agents in organisations using social commitments. Electronic Notes in Theoretical Computer Science, 150, 73-91. Carton, F. (1984). Autonomy behaviour systems of language learning (Systèmes autonomisants d’apprentissage des langues). Paper presented at the workshop “Autonomy and Teaching French” at the 6th World Congress of the International Federation of Professors of French (FIPF), 15-20 July 1984, Québec, Canada. Demazeau, Y. (1995). From interactions to collective behaviour in agent-based systems. Proceedings of the 1st European Conference on Cognitive Science, Saint-Malo, France, pp. 117-132. Hoc, J.-M (1996). Supervision et Contrôle de Processus: la Cognition en Situation Dynamique. (Supervision and control of process : cognition in dynamic situation). Presses Universitaires de Grenoble, Grenoble, France. Inagaki, T., Moray, N., Itoh, M. (1998). Trust self confidence and authority in human-machine systems. Proceedings of the 7th IFAC/IFIP/IFORS/IEA Symposium on Analysis, Design, and Evaluation of Human-Machine Systems, September 16-18 1998, Kyoto, Japan, pp. 431-436. Millot, P. (1998). Concepts and limits for HumanMachine Cooperation. Proceedings of the IEEE SMC CESA'98 Conference, Hammamet, Tunisia, April. Millot, P. and Hoc, J.-M. (1997). Human-machine cooperation: metaphore or possible reality? Proceedings of the 2nd European Conference on Cognitive Science, April 9-11, Manchester, UK, pp.165-174. Millot, P. and Vanderhaegen, F. (2008). Toward cooperatve and human error-tolerant system. Proceedings of the 17th IFAC World Congress, 611 June 2008, Seoul, Korea. Polet, P., F. Vanderhaegen and R. Amalberti (2003). Modelling Border-line tolerated conditions of use (BTCUs) and associated risks. Safety Science, 41, 111-136. Polet, P., Vanderhaegen, F., Wieringa, P.(2002). Theory of safety related violation of system barriers. Cognition Technology & Work, 4, 3, 171-179 Saurin T. A., Formosao, C. T., Cambraia, F. B. (2008). An analysis of construction safety best practices from a cognitive engineering systems perspective. Safety Science, 46, 1169-1183.
However, the feasibility of applying such humanmachine cooperation principles is often limited to the study of static knowledge implemented on the automated decision support tools. Indeed, the cognitive capacity of automated system is usually limited and cannot evolve dynamically. This paper has then proposed an original view on the decisional autonomy control process applied for human-machine systems. It developed the so-called know-how-tolearn in order to manage dynamically the knowledge of the automated tools and to increase the system autonomy. The proposed approach focused toward a feedforward-feedback learning based process that implement the activities Future development will study the feasibility of managing dynamically the automated system autonomy in terms of cognitive capacities to make decisions possible facing know and unknown situations. Such future systems will be able to autolearn from human errors, from interaction with humans, etc. Cooperation principles will be then applied not only for adjusting the human-machine system autonomy by sharing physical resources, resource activities or activity supports between human and machine, but also for increasing the decisional autonomy of the implemented automated decision support tools. Automated tools with such learning capacities will be able to build a model of their users, to identify possible diverted uses and to control the associated risks of such wrong uses. ACKNOWLEDGEMENTS The present research work has been supported by the European Research Group on Human-Machine System in Transportation (GDR E HAMASYT), the International Campus on Safety and Intermodality in Transportation (CISIT), the European Community, the Délégation Régionale à la Recherche et à la Technologie, the Ministère de l'Enseignement Supérieur et de la Recherche, the Région Nord Pas de Calais and the Centre National de la Recherche Scientifique. The authors gratefully acknowledge the support of these institutions.
Sheridan, T. B. (2000). Function allocation algorithm, alchemy or apostasy. International Journal of Human-Machine Studies, 52, 203-216. Vanderhaegen, F. (2004). The Benefit-Cost-Deficit (BCD) model for human analysis and control. Proceedings of the 9th IFAC/IFORS/IEA symposium on Analysis, Design, and Evaluation of Human-Machine Systems, Atlanta, GA, USA, 7-9 September 2004. Vanderhaegen, F., 1997. Multilevel organization design: the case of the air traffic control. Control Engineering Practice, 5, 3, 391-399. Vanderhaegen, F. 1999a. Cooperative system organisation and task allocation: illustration of task allocation in air traffic control. Le Travail Humain, 63, 3, 197-222. Vanderhaegen, F. (1999b). Multilevel allocation modes - Allocator control policies to share tasks between human and computer. System Analysis Modelling Simulation, 35, 191-213. Vanderhaegen, F. (1999c). Toward a model of unreliability to study error prevention supports. Interacting With Computers, 11, 575-595. Vanderhaegen, F. (2003). Analyse et contrôle de l’erreur humaine (Analysis and control of human error). Hermès Science Publications – Lavoisier, Paris, France. Vanderhaegen, F., Chalmé, S., Anceaux, F., Millot, P. (2006). Principles of cooperation and competition - Application to car driver behavior analysis. Cognition Technology & Work, 8, 3, 183-192. Vanderhaegen, F., Zieba, S., Polet, P. (2009). A reinforced iterative formalism to learn from human errors and uncertainty. Engineering Applications and Artificial Intelligence, 22, 654659. Xiabo, Z. and Ohno, K. (2000). Properties of a sequencing problem for a mixed model assembly line with conveyor stoppage. European Journal of Operational Research, 124, 560-570. Zieba, S., Vanderhaegen, F., Polet, P., Debernard, S. (2009). Resilience of human-robot system using ajustable autonomy and human-robot collaborative control. International Journal on Adaptative and Innovative, 1, 13-29.