Affect Control Theory and Social Models of Human-Computer Interaction

One of the most interesting models of emotion is Affect Control Theory (ACT) (Heise, 1987). ACT expresses how social events are construed positing a relation between cognition and emotion. In contrast to psychological theories of social cognition, ACT emphasizes that meanings are culturally shared and deviations from meanings generate arousal that triggers re-appraisals. In this brief article and following to Troyer (2004), an outstanding specialist, I expose some aspects of the theory and how Lisa Troyer applies its concepts to model Human-Computer Interaction (HCI).

ACT looks at individuals as agents seeking consistency across interactions. Humans are categorized into roles with shared expectations regarding actions appropriate for the role. Actions in response to one another are markers of social events. Social events are formed by actors who assume identities, behaviors of actors and objects to whom the actions are directed. For instance, a policeman helping citizen, whose linguistic structure would be "Policeman Helps Citizen". Its meaning is defined in three dimensions: Evaluation (goodness), Potency (powerfulness) and Activity (liveliness). Ratings of each element are called "fundamental sentiments" (for it the word affect). Humans have culturally shared fundamental sentiments and expectations. So, we expect good policemen to behave in right ways. When elements combine in an event, emotion signals the corespondence between the meanings we expect and the actual meanings evoked. Smith-Lovin (1987) introduced impression-formation equations combining ratings of elements to estimate new ratings for inputs combined in events. These equations predict how meanings shift as interaction evolves. The sum of the squared differences between fundamental sentiments of the elements and impressions from the event generates the perceived likelihood of an event.

ACT includes a database of ratings for thousand of elements and modifiers (emotion labels for roles, for instance, "pleasant policeman"). Using the equations and database, ACT predicts the basis for expectations in subsequent interaction. The models and database are combined in the software INTERACT. With this software researchers simulate events and generate testable predictions regarding sequences and redefinitions of events. ACT has focused on Human-Human interaction but Troyer (2004) demonstrates how can be used to model Human-Computer interaction. ACT does not require that computers exhibit emotions, but only that they be able to reason about them, that is, metareasoning.

Troyer ("Affect control theory as a foundation for socially intelligent systems", 2004) designed an experiment with 15 subjects, providing them independent ratings for elements of Human-Computer interaction: "Computer", "Run Analysis", "Provide Output", "Freezes" and "Runtime Error". Correspondent social concepts were "Academic", "Ask about Something", "Educate", "Beg" and "Laughs At". Troyer substituted the correspondent social concepts for HCI terms to simulate events representing Human-Human interaction analogs of HCIs. The simulations explored how meanings shifted when a computer initially behaved as expected, producing unexpected behavior. The simulations showed how different events produced different definitions of the actor eliciting the behavior (grouch/delinquent) and the responses to that actor (scold/avoid). Besides, ACT predicted the emotions of the object receiving the behavior.

ACT may provide an architecture for designing socially intelligent systems.

Physarum machines

A Physarum machine is a programmable amorphous biological computer experimentally implemented in the state of slime mould Physarum polycephalum. It comprises an amorphous yellowish mass with networks of protoplasmic tubes, programmed by spatial configurations of attracting and repelling gradients. It feeds on bacteria, spores and other microbial creatures. When foraging for its food the plasmodium propagates towards sources of food particles, surrounds them, secretes enzymes and digests the food. The plasmodium is considered as a parallel computing similar to existing massive parallel reaction diffusion chemical processors. Adamatzky (University of Bristol) and collaborators demonstrate how to create experimental Physarum machines for general purpose computation: plasmodium can implement the Kolmogorov-Uspensky (KUM) machine, a mathematical machine in which the storage structure is an irregular graph. KUM is defined on a labeled undirected graph with bounded degrees of nodes and bounded number of labels. KUM executes several operations on its storage structure: SELECT an active node (that is, occupied by an active zone) in the storage graph; SPECIFY the node´s neighborhood; MODIFY the active zone by ADDING a new node with the pair of edges, connecting the new node with the active node; DELETE a node with a pair of incident edges; ADD/DELETE the edge between the nodes. A program for KUM establishes how to REPLACE the neighborhood of an active node with a new neighborhood, depending on the labels of edges connected to the active node and the labels of the nodes placed in proximity of the active node. In Physarun machine, a node of the storage structure is represented by a source of nutrients, an edge connecting two nodes is a protoplasmic tubes linkink two sources of nutrients corresponding to the nodes. Finally, an active node is domain of space (which may include nutrient sources) occupied by a propagating pseudopodium. The computation is implemented by several active zones. It uses distributed local sensory behaviours, approximating phenomena observed in Physarum, like foraging for food stimuli, amoebic movement, network formation, network minimisation, surface area minimisation, shuttle streaming, spatially distributed oscillations or oscillation phase shifting. The emergent plasmodium behaviours are represented taking a multi-agent approach or based upon particles. In fact, movement and internal oscillations of the plasmodium reflect the collective behaviour of the particles population. The movement of agents correspons to the flux of sol within the plasmodium. Cohesion of the plasmodium arises due to the agent-agent interactions and movement of the plasmodium is generated by coupling the emergent mass behaviours with chemoattraction to local food source stimuli. Agents both secrete and sense approximations of chemical trails being the population represented on a two-dimensional discrete map. The strength of the projected food sources can be adjusted using parameters and when the plasmodium engulfs a food source the stimulus for diffusion is reduced by the encapsulation. The diffusion gradient corresponds to the quality of the nutrient and substrate of the plasmodium´s environment, and differences in the stimulus strength, stimulus area, affect both the steepness, and propagation distance of the diffusion gradient and affect the growth patterns of the virtual plasmodium.

The Physarum machine by Adamatzky is a very interesting example of a green computer, showing the necessity for simulating simple behaviours, in first place, as motivation for developing more complex simulations. It will be an excellent source of ideas for anyone who is inspired by emerging non-silicon computers.

Pandas: Pandaemonium-Controlled Animats

(Artificial Life, vol. 16, 1, Winter 2010, The MIT Press; cover by Philip Beesley´s Epithelium, an installation at the Siegel Gallery at Pratt Institute of Design in Brooklyn, N.Y., 2008).

On this month I would like to expose an interesting model of cultural transmission developed by Chris Marriott, James Parker and Jörg Denzinger at University of Calgary. It is called PANDAS and simulates the effects of an imitation mechanism on a population of animats (artificial animals) capable of individual ontogenetic learning.

Let be a world inhabited by animats and consisting of a discrete grid that contains besides the following objects: food, water, cave and tree. The world obeys several rules of evolution: (a) no object can occupy the same cell as a tree; (b) all objects are stationary except pands; (c) water, caves, and trees are static and (d) food is depleted when used, and grows over time. For the pandas, we have these rules: (1) pandas can move one cell in one of eight directions (N,NE,E,SE,S,SW,W,NW); (2) pandas obtain food energy, water energy, rest energy and have a health parameter and, in every round, lose a fixed amount of food energy, water energy and rest energy; (3) pandas die if any parameter in (2), drops below cero; (4) a dead panda is removed from the grid and replaced with food; (5) a panda can only mate with another mating panda; (6) a panda can create a new panda by mating or spawning but a panda can only mate or spawn when it has an excess of all energy types; (7) pandas can only interact with objects in the same cell; (8) a panda can fight with another panda, reducing the target panda´s health and (9) a panda´s health automatically recovers, at an energy cost.

A panda has only a one-cell perception range and has a limited internal sense allowing it to monitor its energy levels. All perception is done using input daemons. Daemons are similar to nodes in a network playing a specific role. The input daemons that form the input layer of the panda include 9 daemons for each object in the environment (see food, see water, see cave, see tree). There is one daemon for each of the cells in the perceptual range of the panda. The input layer contains the following set of daemons (food energy, water energy, rest energy, hungry, thirsty, tired). The cognitive cycle of a panda begins with taking input from the environment and activating input daemons, and ends with an action selected. When an action has been selected, it is executed in the system and the cycle begins with the new situation.

The model is very interesting because pandas can engage in three types of adaptation: the genome of a panda encodes both "physical properties" (the panda interacts with its environment) and "mental properties" (the panda maintains its internal organization). To study the effects of the imitation on the populations of pandas, it is considered a group of pandas without imitation drive and unable to perceive directly and a second group of imitating pandas. The authors use the life span of the pandas as an indicator of their success. To do this, they introduce a parameter called the elder age, that is, an arbitrarily selected age such that when a panda reaches this age before dying, is considered successful and the value is used for the measurement of frequency. According to Marriott, Parker and Denzinger, the data of the experiments using this model, show that the median frequencies for imitating pandas are hogher than the frequencies for non-imitating pandas. Obviously, the imitating pandas have the tendency to group together, being particularly strong in newborn pandas. The model supports that mechanisms of cultural transmission can increase the frequency of success in a population but the authors look for extending it to the more sophisticated mechanisms of the "true imitation" because they think that the attribution of higher-order intentionality to artificial agents is something pending nowadays.

Gaze following and mirror neurons

Gaze following is a basic component of the human social interaction and it is a type of attention sharing behaviors. It is also present in a number of other species (for instance, apes) and seems necessary for designing social robots. It can be defined as the ability to look where somebody else is looking. Triesch, Jasso and Deák (2007) have formulated a computational model of gaze following by means of ideas related to the behavior of mirror neurons. The authors emphasize the role of learning processes by means of the interaction with the social environment. In fact, gaze following can be linked to imitation. The link between gaze following and imitation also implies the similarity between the neural basis of gaze following and the neural basis of other imitative behaviors. Triesch and collaborators develop a model that share properties with mirror neurons, that is, neurons implicated in imitation and originally founded in macaque area F5 by Rizzolatti and his team in Parma University. The model predicts the existence of a new class of mirror neurons for looking behaviors that has not been observed experimentally. In this model, an infant and a caregiver interact with a number of visually salient objects. During the process, the infant learns to predict the locations of salient objects based on the looking behavior of the caregiver. There are periods when the caregiver is present and periods when the infant is alone with the objects. When the caregiver is present, the infant and caregiver are in fixed locations facing each other with a separation between them. At any time a random number of objects will be present. Habituation decreases the perceived saliency of an object.

The infant model learns through a reinforcement learning scheme. The learning process tries to optimize the infant´s policy, that is, the way the actor maps sensory states onto different gaze shifts in order to maximize the long-term reward obtained by the infant. Reward is obtained as the saliency of the position to which attention is directed after a gaze shift has been made. At each time step, the caregiver looks at the most salient object, where saliency is mediated by the same habituation mechanism as in the infant´s visual system. The model neurons in the pre-motor layer share many characteristics with classical mirror neurons. A unit in this layer will be active during the execution of a gaze shift to a certain location in space. This is because the probability of performing such a gaze shift is related to the activation of the unit. The units in the layer will be active when the infant observes the caregiver looking in the corresponding direction. Clearly, the neurons in the pre-motor layer can be viewed as mirror neurons because the combination of being active during execution and observation of a motor act is the defining characteristic of mirror neurons.

Following to the authors, this model can be considered a simple associative learning account of a response facilitation but also has implications for the question of whether mirror neurons are innate or whether they acquire their properties through a learning process. For the mirror neurons concerned with grasping, they find plausible that there are situations where observing an agent grasp an object may predict a reward if the same action is attempted. Such situations may be sufficient for the emergence of mirror neurons for grasping. The reviewed model predicts a very close connection between mirror neurons and imitative behaviors.

Dynamic Neural Fields and cooperative robots

Dynamic Neural Fields formalize how neural populations represent the continuous dimensions characterizing movements, perceptual features and cognitive decisions of agents. Neural fields evolve dynamically generating elementary forms of cognition. Many of social activities are based on the ability of individuals to predict the consequences of other´s behavior. We have to interpret actions of our partners in collaborative works.

Erlhagen et al. (2007) try to understand motor intentions for building artificial agents using Dynamic Neural Fields. The authors consider that action understanding relies on the notion that the observer uses its motoric abilities to replicate the observed actions and its effects. They are inspired in activity patterns of mirror neurons in prefrontal cortex that postulate a chain between neurons coding motor acts. Depending on the specific chain of mirror neurons activated by contextual and external cues, the observer will predict (in a probable manner) what the observed agent is going to do. Erlhagen and collaborators represent the activity of neural populations encoding different motor acts and goals by means of Dynamic Neural Fields. The synaptic links between any two populations in the network is established using a Hebbian learning dynamics.

Let think in autonomous robots which interact in the context of a join construction task and let be a simple reaching-grasping-placing scenario. An observing robot R1 has to select a complementary action sequence depending of the inferred action goal of the other robot R2, or partner robot. So, R2 may grasp an object to place it in front of R1 with the intention to hand it over. Neural populations in the action observation layer and the action simulation layer encode motor primitives such as grasping. Such neural populations in the goal layer are associated with the respective chains in the action simulation layer. To model the dynamics of the different neural populations is used a discrete version of a dynamic neural field. Each dynamic field represents a population of 2N neurons which diverge into an excitatory and an inhibitory subpopulation, each of dimension N. The activation of an excitatory and an inhibitory neuron i at time t is governed by a coupled system of differential equations. The firing rate and the shunting term for the excitation, are non-linear functions of sigmoid shape and the interaction strength between any two neurons within the subpopulation is established by fixed synaptic weight functions which dicrease as a function of the distance between the neurons. A Hebbian learning rule for increasing the synaptic efficacy between presynaptic and postsynaptic neurons is given.

For establishing the chains, the authors employ a learning by observation paradigm in which a teacher demonstrates the sequences, each composed of motor primitives. Once the neural population becomes active, the activity propagates to all synaptically coupled populations. But only a population defining a particular chain will reach a suprathreshold activation level.

The simulation of Erlhagen et al. 2007, shows that the neural representations implementing mechanisms like motor simulation and cue integration may emerge as the result of real-time interactions of local neural populations. But, more important is that they speculate that learning by imitation takes place in two steps: first, the links between chain elements are established allowing a fluent execution of particular action sequences. In a second place, similar action sequences may have a different outcome and then, the focus shits towards the links between the goal and the contextual cues. These consequences will be crucial in the cooperative robotics domain.

REFERENCE
(http://www.mct.uminho.pt/erlhagen/pdfs/ErlhagenEtAl_ICDL07_0004.pdf)

Mirror neurons and synchronization between robots

The discovery of "mirror neurons" in the ventral premotor cortex of the macaque monkey by Rizzolatti and colaborators has generated a genuine impact in Neuroscience. In humans it is impossible to registry the neurophysiological activation of simple neurons but disregarding criticisms (see Alison Gopnik-http://www.slate.com/id/2165123/pagenum/all/-), the influence in many fields of the knowledge (study of social relations, robotics, programming, etc.) is enormous. Privileged witness is the recent work of Barakova, Lourens and Yamaguchi. Barakova and Lourens (2008) design a setup for synchronization and turn-taking behaviour in robots. For it, they connect some aspects of the neuroanatomy of the mirror system in humans with an oscillatory dynamics for neural networks.

In brief, the frontal motor areas receive sensory input from the parietal lobe. Another area is situated in the rostral part of the inferior parietal lobule. Both regions form the mirror neuron system. Besides, the posterior sector of the superior temporal sulcus form a core circuit for imitation. Modelling of the superior temporal sulcus area can be reduced to the influence of the inhibitory neurons, projecting the sensory signals to the inferior parietal lobule, area which is associated with multisensory integration. Each robot has 8 range sensors projecting to the sensory integration area that resembles the functionality of the joined temporal sulcus-inferior parietal areas. The two wheels of the robot project to the sensorimotor integration area, resembling the function of the ventral premotor cortex. Self-organization of rhythmic activity of this system can be simulated by means of endogenous oscillators an so, the change of rate of the phase with the time, is the cycle of the limit cycle oscilation, being the phase periodic over the range.

The experimental setting conceived by Barakova and Lourens presents, in the first place, robots that are taking the role of the follower, in order to establish "mirroring" couplings between the nets that simulate the inferior parietal lobule and premotor ventral areas. Hebbian connections between these nets are modelled such that the interaction behaviour is reflected by the average activation values of unikts over a certain time interval. So, the robot playing the role as a follower, tends to synchronize its motion direction with the motion direction of the leading robot. In the second place, the experiment shows the emergence of turn taking between the robots. The role of the robot, being follower or leader depends on which robot is within the visual field of its partner. The emergent turn taking is expressed by symmetry breaking process after a period of synchronization: the leading robot can become a follower and later the lead can be taken over by it.

Inspired by the mirror system in human beings, Barakova and Lourens simulate very interesting interaction behaviours of following and turn taking such that the mirroring functionality is obtained by means of the selforganization of synchronized neural firing in two robots that share perceptual space. No doubt that many extensions of this work remain to develop and promise great advances in the area of robotics.

The e-pucks and social cooperation

The e-pucks are mobile robots developed at the École Polytechnique Fédérale de Lausanne (EPFL). The designers (see Mondada et al., The e-puck, a Robot Designed for Education in Engineering, 2009) looked for the miniaturization of a complex system combining desktop size and flexibility. These robots have sensors in different modalities (distances to objects by means of eight infrared proximity sensors, accelerometer, color camera, microphones), actuators with different actions on the environment, wired and wireless communication devices and two types of processors (general purpose and DSP). Although they were conceived for education in Engineering, they have demonstrated to be very useful for experimentation in Artificial Intelligence.

Social learning is the capability of an organism to learn by observing the behavior of a conspecific. Evolutionary robotics is a methodological tool to design robots´controllers. Recently, Miglino, Ponticorvo and Donetto (2008)

(http://www.nac.unina.it/images/pub/cooperationincorvid.pdf)

have used an evolutionary algorithm for the neural control of e-puck robots which imitate the cooperation in corvids to obtain a reward (food), which is clearly visible, but not directly reachable. The dyad gets the reward if the two tips of a string are pulled at the same time.

Two e-puck robots are situated in an environment consisting of a square arena and of a corridor both surrounded by walls. The authors use an evolutionary algorithm to set the weights of the robots´neural controller. The initial population consists of 100 randomly generated genotypes that encode the connection weights of 100 corresponding neural networks. Each genotype is translated into 2 identical neural controllers which are embodied in 2 e-pucks. The 20 best genotypes of each generation are reproduced by generating 5 copies each, with 2% of their bits replaced with a new randomly selected value. The evolutionary process is iterated 1000 times and the experiment is replicated 20 times each consisting of 4 trials with 4 different starting positions in the corners of the room. Cooperation between e-pucks is regulated by social interaction with communication as a medium. The emergence of communication leads to a coordinated cooperation behavior similar to cooperation observed in corvids.

Nowadays artificial evolution is employed to build neural mechanisms that control the behavior of learning robots. Mechanisms for social learning in organisms can be successfully simulated in an integrated neural network architecture. E-pucks are an excellent tool for simulating social learning in organisms requiring cognitive mechanisms. But future work is needed, in particular to allow the robots to autonomously decide when to use social learning strategies or individual strategies.

Goran Trajkovski on imitation to modeling agents societies

In this brief review we analyse some ideas about simulation of imitation by means of artificial agents. Dr. Trajkovski is Director of Cognitive Agency and Robotics Lab in Towson University (USA). He is specialized in Cognitive Engineering and has published books like "An imitation-based approach to modeling homogenous agents and societies" or the recents "Handbook of Research on Agent-Based Societies: Social and Cultural Interactions" or "Handbook of Research on Computational Arts and Creative Informatics". Trajkovski introduces agents capable of performing four elementary actions (forward, backward, left, and right) and of noticing 10 different percepts. Each agent is equipped with one food sensor and has one hunger drive. When the hunger drive is activated for the first time, the agent performs random walk during which expectancies are stored in the associative memory. When food is sensed, expectancy emotional context is set to a positive value. Every time in the future the hunger drive is activated, the agent uses the context values of the expectancies to direct its actions. It chooses the action that will lead to expectancy with maximum context value.

Agents inhabit a two-dimensional world surrounded by walls and populated with obstacles. Sensing another agents takes the agent into its imitating mode. The agents have begun to inhabit the environment at different times. They are also being born in different places in the environment. While inhabiting the environment, they explore different portions of the environment that may be very different, or perceptually similar. In environments inhabited by heterogeneous agents, the fundamental problem is the problem of interagent communication. According to Trajkovski, an example of interagent communication is the phenomenon of multilingual agents that can serve as translators. The author proposes an enactivist (Varela, Thompson and Rosch, 1991) representation model, based on the treatment of agents as dynamical systems. The agent during the interaction with the environment generates its inputs and makes a mapping from the continuous domain of the inputs to the discrete domain of the percepts. The sequence of these percepts would reflect the structure of the environment. The basic idea is to divide up the set of possible states into a finite number of pieces and keep track of which piece the state of the system lays in at every iteration. Each piece is associated with a symbol, and in this way the evolution of the system is described by an infinite sequence of symbols. The agent generates symbols or percepts.

Trajkovski shows that imitation is far from a trivial phenomenon and that humans are wired for imitation by means of the research in multi-agent systems. He gives a solid and fertile attempt to establish a mechanism of interagent communication in the multi-agent environment using classical algebraic theories and fuzzy algebraic structures. His contributions are very relevant for the social cognition, mixing studies about animal intelligence (Thorndike) with studies about imitation in humans.

A New Computational Model of Social Learning

We describe a very interesting computational model of social-learning mechanisms proposed by Lopes, Melo, Kenward and Santos-Victor (2009). The authors distinguish between imitation ("adhere to inferred intention, replicate observed actions and effect") and emulation ("replicate observed effect") taking into account the agent´s preferences for different actions and the information available from the demonstration. There is a module addressing the baseline preferences of the agent, evaluating actions in terms of energy consumption. With this module is associated the utility function Qb ranking possible action sequences according to their overall energy consumption. A utility function Qe evaluates the actions in terms of their probability of reproducing the observed result/effect. For the intention replicating module, the utility function Qi assumes the demonstrator is goal-oriented. The module operates by considering all the possible goals in the current system, calculating for each one the relative probability that it would give rise to the demonstrated behaviour, and choosing the one that maximises the probability. In cases that are equally likely to produce the observed demonstration, goals with tied probability are ranked randomly which leads to stochasticity in the final behaviour.

The model of Lopes and collaborators can replicate the tendency to interpret and reproduce observed actions in terms of the inferred goals of the action in line with the experiments of Malinda Carpenter and colleagues (2005): a demonstrator moved a toy mouse across a table from one point to another; in one condition, the final point of the move was inside a little house, and in the other condition, no house was present: infants showed a much grater tendency to replicate the specific mouse moving action observed when there was no house to move the mouse into. The results of the simulation reproduce the findings of Carpenter et al. (2005), confirming the standard interpretation of the experiment: the infant infers what the demonstrator´s intention was.

Other simulation replicates the famous experiment of Meltzoff (1988) in which 14-month olds were exposed to a demonstrator who performed unusual actions on objects (there was a box with a panel that lit up when the demonstrator touched it with his forehead, and most infants copied the use of the forehead rather than using their hand): the infants reproduced the actions with a delay of a week. According to Lopes and collaborators, the simulation confirms the imitation in terms of the inferred intention and of sensitivity to the constraints on the demonstrator.

To investigate what happens when the learner does not have complete knowledge of the world dynamics, they model a type of experiment based on Horner and Whiten (2005) experiment in which presented preschoolers and chimpanzees with two identical boxes, one opaque and one transparent. The demonstration consisted of inserting a stick into a hole on the front of the box, with the latter step generating a reward. The insertion of the stick into the top hole was unnecessary in order to obtain the reward, but the causal physical relations were only visible with the transparent box. The results showed that 3 and 4-year-old children imitated both actions no matter whether they had observed and were tested on the transparent or opaque box, but chimpanzees were more able to switch between emulation and imitation after having observed demonstrations with a transparent box and reduced tendency to insert the stick into the ineffective hole.

The simulation designed by Lopes, Melo, Kenward and Santos-Victor replicate the results from both children and chimps when the weight of the intention replicating module was increased, confirming that the difference occurs because chimps are primarily motivated to select the most efficient method they know to achieve the end effect.

Following the taxonomy proposed by Call and Carpenter (2002), Lopes et al. (2009) build a unifying mathematical model of types of social influence on behaviour, mainly imitation and emulation, concluding that a switch between imitation and emulation might be triggered by changing the value (to the learning) of the social interaction or of the effect. So, the greater utilization of imitation by children might be explained by a stronger focus on others´intentions, mediated by social cues.

Imitation, Maslow´s Pyramid and Multi-Agent Systems

We review an interesting article published by Le Guen Herve and Moga Sorin in Proceedings of International Joint Conference on Neural Networks (2009).

The authors study the influence of imitation within a population of artificial agents following Maslow´s Pyramid of needs. Imitation is the possible origin of communication and social learning. The imitative abilities include many types like facial imitation, perception by the infant that he or she is being imitated and empathy. The interest in empathy induces the study of emotions which Herve and Sorin link to the analysis of social needs as modeled by the Maslow´s Pyramid of motivations. Although the influence of each particular need varies from one person to another the principle is there are two main classes of needs. The most important needs are called primary needs (eating, breathing, etc.) while the second class of needs constitute the social aspect of human motivation and includes the need for esteem. This need is bidirectional and has been interpreted by the authors as the need to imitate and the need to be imitated. As the fact of being imitated is perceived by the agent concerned and as imitation is likely to generate empathic satisfaction, the authors model them as expected rewards. Based on a model of an autonomous robot with goal-oriented navigation and imitation capabilities, the robot goals are derived from internal variables that have to be maintained in a comfort area. The values of these variables decrease in time. The robot population´s task is to explore an unknown environment and to localize sources corresponding to its needs. Its survival will depend upon the satisfaction of these needs by discovering the different locations of the sources. The behaviors are obstacle avoidance, goal-oriented navigation, imitation and exploration. The emotional signature expresses the current state and the expected state of the agent: an agent in its comfort area displays a neutral signature whereas an internal variable below a certain threshold induces pain. That pain will cause a potential empathic response that is likely to incite another agent to move toward a known source. Besides a motivated agent is likely to provoke an attractive empathic response: keeping a motivated agent in its own field of perception becomes a source of motivation. It leads to the selection of the imitation target according to its motivational state.

The results of the simulation with Maslow agents showed an enhancement of the global survival rate even with a very small population wherein communications are not frequent.

The authors have proposed a holistic approach to the implementation of imitation in autonomous agents.

We recommend this excellent article because its ideas permit to build a simple and scalable model of agent useful in applications such as networks or swarm piloting.