Towards an Emotional Decision-Making Mickaël Camus1 2 , Alain Cardon2 1: L.E.R.I.A. {Epitech.} 24 rue Pasteur 94270 Le Kremlin Bicêtre, France 2: LIP6 UMR 7606 Paris VI, UPMC 4 Place Jussieu, 75252 Paris Cedex {Mickael.Camus,Alain.Cardon}@lip6.fr

Abstract Industrial need new technologies to make evolved methods, strategy and capacity. Decision-making plays an important role in the case of the robot evolving in an instable environment, and enables us to decreasing the human factor in the control system. This paper presents a model based on brain construction to give a restricted emotion to a machine in order to control a multitude of entities. Emotion permits to make a rapid decision in a hostile environment. The method used to build this system is based on Massive MultiAgent System (Massive MAS). It enables us to have a vast number of entities with an asynchronous communication.The morphologic aspect is used to observe the agents behavior with the aim to generate a restricted emotion in order to make an action plan.

1

Introduction

New technologies help industrial to develop knowledge in communication, exchange, security, cryptography and others. Data and decisions are critical, that is the reason why all computing tasks are managed by a human factor. A good example is the Unmanned Aerial Vehicles (UAV) to search information. This technology is controlled by humans, but the dangerous task is managed by the machine. It is an interesting technology but it is hard to be used in an hostile environment. It is a reason to give a more important role to a computer. This decision allows to mimic a human in a decision-making process. There is important needs in computer science to succeed in this ambitious project because dataflow is very important in an unstable environment in constant evolution. A good solution would be a control system software based on human emotions [1] for the decision making, with a simulation section for the training and the composition of autonomous. Some work proved that emotions have a link with human decisionmaking such as Bechara’s work presented in [2] or also Lerner’s work presented in [3]. A software with these features is in development, its codename isPALOMA. It is developed with the oz/mozart system [4] [5] to have a maximal concurrency and a high message passing. This paper focuses on the decision-making model.

1

2

2

Aims

The human brain process different pulse and allow human-use of nervous system. Pulse have a role in the knowledge processing which is saved in the gray matter, and it is proved that emotion is a pulse and has also a role in the decisonmaking. Matter, pulse and physiology are the base of the brain operation. This unit is adaptative, allows to generate decisons, emotions, create thoughts and ideas. The goal of this project section is to build an adaptative system mirroring the emotional and inentional human behavior as suggested and developed by Cardon [1] to make decisions in an unstable environment with a multiagent system.

3

Project

The project PALOMA is developed to simulate and build an autonomous system. It provides for the construction of an environment for a particular problematic with coherence, naturalness and reality. This tool is described by Camus and El Kadhi in [6]. As a part of the global PALOMA project, this paper details the development of the decision-making level (DML) based on the brain construction for a Sony Aibo. In this DML, We rely on a particular approach based on the mirroring of the human nervous system as described by the neuroscience research community [7]. The goal of this section is to build a system allowing an intentional behavior based on emotions.

4

Machine representation

In this work, a machine is represented as an entity with two levels: hardware and software. These two levels cooperate and evolve together. The hardware is composed of multiple sensors and effectors. The software is, in fact, the brain of the hardware. It is not embedded in the machine, but is installed on a workstation where several calculus of the software can be distributed on a network, such as in the figure 1. In an Aibo, we have engines for legs, a microphone, a video camera associated with different sensory sensors and distance sensor to interpret a scene. Each sensor and effector has a role and measures a specific action in an environment. The “brain” component for the Aibo application is, for us, a multiagents system gathered in a polymorphic geometrical form that represents the emotional behavior of the entity during a decision-making event.

5 Multiagent system architecture and aspectual agent MultiAgents System enables us to mimic the complex mechanism of the generation of emotions [8]. To reach this goal, different roles are defined. They allow the data process with importance and quality priority. The term used here is “agen-

3

User Interface

Brain

Data Sensors

Workstation Environment Wifi Communication

Robot

Scene

Data Effectors

Figure 1: Workstation and environment

tification" in french, the effectivness system depends on this action. Definition : a role is related to a knowledge base. It is for example, the recognition of a specific object saved in the knowledge base in the environment. Agent structure is common to all the MAS, only their role differs. A role is a domain of knowledge, there is different group of roles in the system. This structure is presented in figure 2. It is simple because the important points are the system organization, the roles, the agents number and not the agent complexity. An agent is composed of a knowledge base with intentions and goals, with a communication API and an inference engine. Definition : a goal is a search action to find values which are in the domain of the knowledge base. Definition : an intention is a list of actions to succeed a goal. Knowledge base is crucial, it directs an agent towards a specific role. For example, in a UAV, an agent can be interested in the obstacle detection and not in the threats on the ground. These threats will have a particular interest for other agents. A central element in this architecture is the agent state. The different states are represented by a dynamic graph, with a history, and the possible states. The actual state is transmited towards agents with different roles.This state is conform to the knowledge base. This information transit is very important, it is the core of the system. The communication API must be powered to decrease the latency time between requests and answers.

4

Knowledge Base Intentions and Goals In Inference

Com API

Engine

Out

State

Figure 2: Architecture of the aspectual agent

The agent knowledge allows to have a state on process actions. The processing of actions are called "intentions”. After a data processing, an agent has intentions to reach a specific goal. This system is presented in Figure 3 where we can see an intention stack and a list of goals. These two elements are synchronized and communicate directly with the inference engine to treat the base.

6

Brain features

Let us, now, specify the capacities of the “brain” system. To evolve cleverly in an instable environment, a robot must be able to collect data, analyze, makedecisions and then act according to its goals. This is the natural human behavior such as presented in [9] by Kolb and Whishaw. Our approach to decision-making in an hostile environment allows us to divide these capacities in five crucial levels: 1. Represent a contextual situation. 2. Direct the attention on particular elements (objects or actions). 3. Feel emotions based on these elements. 4. Build behavior action plans. 5. React to the action feedback of an object. These levels are developed step by step with a multiagent system vision and are detailed in the following sections.

6.1 Description The first level is a scene representation. Aibo sensors, such as sensory sensor, distance sensor and the video camera, allow to process environment data to give a contextual position to the Aibo. The goal of the camera is to recognize some simple forms such a door, a wall, a chair and others. The camera data are processed

5

Intentions and Goals Intentions Stack

Goals

Sync

Communication

Inference Engine

Figure 3: Goals ans intentions

by a neural network embedded in any agents of the system in order to build a little vision ontology. This vision ontology is associated with the sensors ontology and is distributed in the multiagent system with different links. The second level associates the goals of the robot with its environmental knowledge. The goal is to give priority to some objects or actions in the environment. Let us take an example: if the goal of the Aibo is to be near its owner all the time, the processing priority is the movement of the owner for the action, and, for example, the shoes of the master for the object. This method allows us to select data in the information flux to increase the system treatment. The third level works on the multiagent system morphology to detect particular geometrical forms stable in a timeline. This form is analyzed by a method presented by Campagne in [10]. It allows us to recognize and classify the different geometrical forms as an emotion generated by the robot during its evolution in the environment and during the recognition of objects and actions. The fourth level creates a relationship between the cognition and the action. The word “behavior” is used for the “reactions” of the robot to the environment (the robot uses its effectors to act on the environment). For a cognition degree, there is a succession of actions on different effectors. More the cognition degree is high, more the list of actions is specified. The fifth level is a continual and infinite bidirectional interaction and adaptation between the environment and the robot behavior. For each action, there exists feedback, a relationship between sensors and effectors. The attention of the robot evolves with the actions processed in the environment. Cognition and action are treated in parallel in the multiagent system.

6

6.2 Organisation The organisation of the multiagent system is based on the “aspectual” agent type presented by Cardon in [11]. It is a multi-level representation. In the case of an Unmanned Aerial Vehicle (UAV) simulation for example, UAV is not limited at an agent, but is represented by a multiagent system. It represents a situation by features (called structuring organisation which is also presented in [11] and [1]), interpretes a scene to make decisions, and finds the adapted moment to act on the environment with a continual utilisation of the morphologic control. The system evolved in the environment with a systemic loop: sensors −→ representation −→ effectors −→ sensors. An emotion is generated with an emergence on several knowledge. Emotion is not a simple variable with an applicative field such as fuzzy logic, but rather an agent processus with a particular activity form in the structuring organization of the multiagent system. There is a strong analogy on the operating mode and the data exchange between the meso-limbic and the neocortex of the human brain. A human or any other entity evolving in an environment treats all information simultaneously. The DML presented in this paper obeys the same constraints and suggests a parallel information treatment and communication. This theory is exposed by Kolb and Whishaw in [9] where authors present the different links between brain and behavior in general. Another major fact is the representation of links between sensors and effectors. For our case, this representation ensures the exchange between hardware and software. We suggest gathering our system components (sensors and effectors) according to a simple record based on sensor and effector capacity. This method allows us to add or delete sensors or effectors dynamically to increase or decrease the global capacity of the system.

6.3 Message passing and semantic Exchange data and organisation can be processed with a communication between agents. This communication is specific for each role type. Oz language allows us to specify a particular design to manage, create, or delete messages linked to a role. With this exchange, several agents have acointance with others and we can observe the communication between agents to note an emergence on one or several roles [11]. The communication is crucial. We used semantic in thread message passing to mirror pulse, physiology and matter used in the brain. We can have questions on the semantic using such as: how a word is represented in the brain? where this word is saved? how uses this word to compose a sentence? Several research have proved that the brain is a continue and parallel serial image [12]. This image is view in tri-dimension, it is a four-dimension view of the brain. This processing proved that a word, a knowledge, is not saved in a particular region of the brain. All the knowledge is distributed in the brain and we used pulse,

7

Mutiplexer

Agents n messages

Managed

Agents n messages

Messages n

n

Figure 4: Communication system between agents

physiology and the matter to create a specific geometrical form adapted to a context in the environment [13] and communicated at ohter person with the language. More a person has vocabulary, more it could explained it. Here, semantic is used to build a definition, an abstraction of an emergence. This emergence is a strong link between different knowledge distributed in the system. This knowledge is represented by a tri-dimensional graph1 with different distance between words2 .

7 Results Severals levels in the developpement have been checked such the control system based on the morphology, the asynchronous communication between agents and the systemic loop. The theory of the control system has been validated by the defense of a Philosophia Doctor thesis by Campagne. The title of his thesis is in english: “Multiagent System and Morphology” [14]. This theory was been applicated with the language Oz. The asynchronous communication between agents has been developped with the concurency feature of the Oz/Mozart system. This feature is composed of “thread” and “message passing”, tow tools directly include in the language. There is different type of message in the organisation of agents. A mutiplexer has been developped to manage all messages in the developped system. With this tool, it is easy to send a message towards a specific agent or a specific group. Figure 4 shows the multiplexer and its role in the communication system. 1 This

graph is a parallel view of the general communication between agents in the system. word can be an object such as a ball, a keybord, or a verb such as “Play”, “Work” and other. It is possible to have an adjective such as “Beautiful” or “Tall” 2A

8

Figure 5: Communication between agents

The systemic loop allows us to treat data of different Aibo sensors to build a representation of the scene, generate a restricted emotion, make a decison and applied this decision with the Aibo effectors and evaluated the result with the processing of the return data. This result show us a specific organisation, a knowledge emergence such as in the figure 5. Currently, the movements of the Aibo are limited besause the ontology is tiny at this time. Now, the work is to increase the ontology and linked this ontology to more capacities of the robot. Tools are developed for the DML, such as a user interface with an “Opengl” representation of the agents (Figure 5), to create different experiences and follow the evolution of the geometrical form generated by the structuring organization: an artificial mental represention of the emotion such as shown in the figure 6. It is possible to give different parameters to the application to build the system such as the sensors, effectors and agents number. The different links between the agents, or the control method for the communication in the multiagent system.

8

Conclusion

Decision-making without human factor is crucial for an autonomous entity such as a robot, an unmanned vehicle, or a server in an international compagny. To succeed in autonomous system, it is interressant to abord different point of view, and one of these view is to create a system which mimic the human nervous system. Mirroring the human brain is also the aim of the artificial intelligence, may be that the existent technology allow us to create this system, it is a one of the major goal of this project. Multiagent system allows us to mimic the intentionality production in the brain. It is a specific organisation of the system with an emergence on several knowledge recognize in the environment to build a representation of a scene. This method is

9

Figure 6: Geometrical form

possible with a distribution of the knowledge base in the multiagent system. This emergence highlights a part of the knowledge of the system. If these knowledges match with the goal of the system, a decision can be made, an emotion is highlighted. This decision is not optimal, there is a continue evaluation to proved the quality of the decision and direct the system on a part of its knowledge to succeed a goal if it is necessary. All the system is processed such as a systemic loop evolving in an instable environment. Knowledge are updated in the time line, this event increase the capacities of the system. The project is developed with the system Oz/mozart [5]. Oz is a multi-paradigm language with scripting programming, object programming, logic programming and constraints programming. It allows us to use different paradigms such as the concurrency to develop a multiagent system with asynchronous communication using the message passing, or the constraints programming to create different action plans. It will be interessant to try a direct learning to use the different sensors and effectors on a robot to transformed a hardware system to a polymorphic hardware system.

References [1] Cardon, A.: Modéliser et concevoir une machine pensante. Vuibert (2004) [2] A. Bechara, H.D., Damasio, A.: Emotion, decision making and the orbitofrontal cortex. Cerebral Crotex 10 (2000)

10

[3] Lerner, J., Keltner, D.: Beyond valence: Toward a model of emotion-specific influences on judgment and choice. Cognition an Emotion 14 (2000) 473– 493 [4] Roy, P.V.: Oz/mozart environment (2002) [5] Roy, P., Haridi, S.: Concepts, Techniques, and Models of Computer Programming. MIT Press (2004) [6] Camus, M., El-Kadhi, N.: Generic simultor environment for realistic simulation - autonomous entity proof and emotion in decision making. Journal of Systemics, Cybernetics and Informatics 2 (2004) [7] D. Purves G. J. Augustine D. Fitzpatrick L. C. Katz, A.S.L.J.O.M., Williams, M.S.: Neurosciences. De Boeck Universite (2003) [8] Cardon, A.: Conscience artificielle et systèmes adaptatifs. Eyrolles (2000) [9] Kolb, N., Whishaw, I.: Cerveau et Comportement. De Boeck Universite (2001) [10] Campagne, J., Cardon, A.: Using morphology to analyse and steer large multi-agents systems at runtime. In: Selmas 2004 IEEE, Edinburg, Scotland. (2004) [11] Cardon, A.: Design and behavior of a massive organization of agents. Design of Intelligent Multi-Agent Systems, Human-Centredness, Architectures, Learning and Adaptation 162 (2004) 133–190 [12] Channouf, A., Rouan, G.: Emotions et Cognitions. De Boeck Universite (2002) [13] Thom, R.: Modèles mathématiques de la morphogenèse. Christian Bourgois (1989) [14] Campagne, J.: Morphologie et système multi-agent. PhD thesis, Université Pierre et Marie Curie (2005) draft.