From: AAAI-98 Proceedings. Copyright © 1998, AAAI (www.aaai.org). All rights reserved. Emotion-based Agents* Rodrigo M. M. Ventura and Carlos A. Pinto-Ferreira Instituto de Sistemase Robdtica,Instituto SuperiorTe’cnico Rua Rovisco Pais, 1 1096 1---1 ishoa- (Index. --_---a--, -Portnod --_I DI- (yoda,cpf}@isr.ist.utl.pt . To survive in a dynamic and rich environment. human beings have to p>ocessvery complex stimuli in real time. Whatever artificial system satisfying the challengeof achieving a similar performancein complex data handling, consideredspecies. When an unknown object appearsin the scene,its perceptual image is extractedand a first estimate of the vector of desirability, as well as the correspondingdecision making in owht -. 51--. to .- incomorate --.---=-..__ mechanisms -- .-- _._..--__A. nnecificallv -r- _.__----, the &gfi cnncejv\red_ tc, perform efficiently. We hypothesizethat such efficiency orientedsystemsprocess stimuli - simultaneously- under two different perspectives:a cognitive, elaborative- which allows them to understandwhat is happeningand what they know about the world, and a perceptual, immediate- which permits them to react quickly and decide adequatelyin circumstancesdemanding urgentaction. Hence,from the very samecomplex stimulus, two sets of facets are extracted: one, mostly directed to recognition and reasoningpurposes,and another, aiming at assigningdegreesof threat, danger,pleasure,and so on, to the current situation, constructingwhat we call a vector of desirability. For instance,when faced with the image of a moving object, the cognitive processorprovides elementsto recognition (is it a lion or a rabbit?), whereasthe perceptualprocessordelivers an assessmentof the prevailing color, moving speed,dimension, and other relevant featuresfound in the scene (is it a huge object with a particular color - a predator,or a little quick moving object - a prey?). These characteristicscomposea “perceptualimage”which serves two purposes:on the one hand, it allows a rough evaluation of the situation and the correspondingdecisionmaking. On the other, it helps the searchwhich underliesthe processof recognition: insteadof comparingthe “cognitive image”under processingwith all the elementsstored in memory, the searchis bound to thoseobjects sharingthe sameperceptual image. To reach this desideratum,“cognitive images”and “perceptualimages”extractedfrom the samesourcestimulus should be associatedand memorizedin such a way that the latter indexesthe former. This kind of system should be bootstrappedby the incorporation of built-in associations.In fact, there should exist some stimuli which are essential, innate: for instance,animals faced with their preys or predatorsdecide either to attack or run away as a function of the vector of desirability a perceptualimage suggests.This assignmentdependson the ‘Copyright(c) 1998,AmericanAssociationfor Artificial Intelligence(www.aaai.org). All rightsreserved. 1204 StudentAbstracts tcmA is y----.z..--. n~rfmtnd nn V.. thP .-A” nther v.1.w. had *...A.-, RQ thP ~-~omt .a” S&A_U~““C doesnot find the correspondingcognitive image in memory (becausethe object is unknown), a new cognitive-perceptual associationis established.This seemsto be the mechanism underlying the Pavlovianreflex - as it lies on top of builtin associations.This also seemsto be the basic mechanism responsiblefor the assignmentof meaningto objects. O..>, _ mcolporanng . ~... -<. mu *l . a’oubie processing and knowi3yJrernb edge representation mechanism, indexing and storing two representations of the same object together with a vector of desirability, are here dejined as emotion-basedagents. This model is corroboratedby the work of several researchersin the field of neuroscience. Namely, in Papez circuit theory the functions performed by the hypothalamus and cingulate cortex (in the human brain) can be identified with the cognitive and perceptualprocessingdiscussed above jieijoux i996j. And according to Damasio’s ideas, emotionsare essentialto human decision making (Damasio 1994). Severalarchitectureshavebeenproposedto deal with real time processingand decisionmaking; in all of them it is possible to identify “cognitive” and “perceptual”levels. However theselevels work independentlyand, in certain cases, one blocks the other. What makes this proposal different and appealingis the intertwining of both, and the consequent iearning capabiiitieswhich it aiiows. An architecturebasedon theseconceptswas implemented and the results are encouraging: not only does the agent learn new associations,but it also recognizesobjects efficiently anddecidescorrectly when faced with new, unknown situations. And interestingly, what can be consideredas an emotionalbehavior,was observed. We are afraid that, to reach intelligent behavior, we will end up implementingsome of the (apparently)most stupid things humanbeingsexhibit, including bad temper. References Damasio,A. R. 1994. Descartes’ Error: Emotion, Reason and the Human Brain. Picador. LeDoux, J. 1996. The Emotional Brain. Simon & Schuster.