Mensagens da Lista de Discussão Ciência Cognitiva

From: rolimms@ig.com.br <rolimms@ig.com.br>
> A psicologia aponta para o principio do prazer. Mario
> aprendeu que estudando para a prova de matematica, obtem
> bom desempenho na avaliacao, obtendo bom desempenho, sua
> mae sorri para ele e seu pai lhe chama de campeao. Alem
> disso, a professora lhe elogia publicamente pelo seu
> excelente aproveitamento.

Esta é realmente uma forma de modelar, mas não parece
ser a única que vale a pena ter em mente. Há certamente
algo relacionado à auto-imagem. O rapaz estuda matemática
porque haveria dano à sua auto-imagem se ele não o fizesse.
Então, a motivação primária de seu estudo poderia não ser
apenas o prazer de ser elogiado, mas também o medo de ser
criticado.

From: Marcelo Bicalho <mbicalho@planetarium.com.br>
> Isto parece simples de ser realizado dentro do esquema "somatório
> de tendências". Cada decisão tomada teria um feedback na forma de
> uma lista de valores positivos ou negativos a serem acrescentados
> em sua "lista de tendências" Assim, se ele joga cartas e ganha,
> na proxima vez sua tendência a jogar em qualquer circunstância
> seria acrescida de + 30, por exemplo (ele estaria "mais confiante"),
> e o inverso no caso de perder.

Esta também é uma idéia razoável, mas precisa ser vista em um
contexto mais complexo. E se o que ocorre não é exatamente uma
soma, mas uma competição? Acredito que muitas vezes o resultado
final de uma decisão (um comportamento) é obtido por uma luta
feroz de diversas "frentes emocionais" competindo entre si.

Há várias vezes em que nos sentimos "pouco à vontade" ao fazer
algo, mas acabamos fazendo a coisa por alguma razão mais forte.

Outro ponto importante a ter em mente é o fato de que grande
parte de nossas razões emocionais parece estar escondida de nossa
cognição consciente. Isto é muito ruim para todos os que tentam
modelar emoções.

A propósito, uma boa dica de textos muito interessantes sobre
modelagem de emoções em computadores é o seguinte:

Prof. Aaron Sloman
School of Computer Science, University of Birmingham
http://www.cs.bham.ac.uk/~axs/

O Prof. Sloman pesquisa isto há pelo menos 2 ou 3 décadas e ele
(e sua equipe) já desenvolveram muitos trabalhos interessantes
(procure pelo Cognition and Affect Project).

Veja em especial a tese Ph.D. de Ian Wright "Emotional Agents",
onde boa parte dessa arquitetura é explicada com detalhes
(a tese é interessante também porque faz um apanhado histórico
de outras arquiteturas). Em 1998 fiz um estudo dessa tese e
preparei um sumário, que copio abaixo. Serve para ter uma
idéia bem rápida dos assuntos tratados por Wright.

--------------------------------------------------------------
Title: Emotional Agents
Ph.D. Thesis (267 pages)
Author: Ian Paul Wright
Cognitive Science Research Centre, University of Birmingham
URL: http://www.cs.bham.ac.uk/~ipw/thesis.ps.Z
Feb. 1997
Summary prepared by: Sergio Navega

This thesis presents a detailed account of several theories of emotion,
besides concluding with the contribution of the author. A theory of
mental pain and pleasure is presented, besides some interesting points
such as that “consciousness” is irrelevant to the understanding of
emotions. Other interesting point mentioned is the concept of
“perturbance”: the effect of emotions on the loss of control of the
thought processes.

The extensive review of similar theories starts with considerations
about Herbert Simon's “interrupt theory” of emotions. The basic point
of this theory is the idea that human behavior have multiple goals and
that motivation controls attention given to a specific goal. The theory
also considers the importance of interruptions to the displacement of
the current goal. An event in the environment, a recalled memory or a
new motivation are factors that can affect the current goal (these are
also known as percepts, thoughts and motivators). As an example, the
appearance of a frightening predator can provoke the interruption of
the current goal of an animal and the establishment of a new one
(fight/flight).

Aaron Sloman's AFP (Attention Filter Penetration) is an extension of
Simon's theory. Among the new architectural concepts introduced, it's
cited the “variable threshold interrupt filter”, which regulates the
ability to disturb of the motivators, thoughts or percepts. Other
important aspect of AFP is the introduction of the “perturbance”
concept, which is responsible to the partial or total loss of control
of attention. This is a characteristic feature of the states we
call “emotional”.

The thesis continues with a chapter about Agent Architectures, in
which some typical agents are presented. Deliberative architectures use
deductive reasoning to solve problems. Examples of such architectures
are STRIPS, NOAH, NONLIN (Classical planners) and GPS, SOAR,
ACT-R (cognitive models).

Reactive architectures are based on behaviour-based approaches,
which avoid the assumptions of unchanging, static worlds made by
deliberative architectures. They are composed of reactive modules
intimately tied to the sensing and acting apparatus of the agent. Each
module is responsible for a particular kind of behaviour. Cited
examples are the work of R. Brooks and Agres & Chapman.

Finally, the Reactive-Deliberative model is mentioned, which try to
combine the best parts of the preceding architectures. Examples cited
include Firby's RAP, Ferguson's TouringMachine, Pryor's PARETO,
Benson and Nilsson's TRP (Teleo-reactive program) and Georgeff's
PRS (Procedural Reasoning System).

Aaron Sloman and Luc Beaudoin's Cognition and Affect project is
succinctly described in their intention to provide human-like autonomy.
A detailed diagram of an architecture for motivated agents is
presented, together with an explanation of each of its modules such as:
perceptual mechanisms, database of “beliefs”, collection of motivators,
variable threshold attention filter and others.

An entire chapter is devoted to the MINDER1 implementation, a
toolkit that was conceived to allow the prototyping of agent designs
under a microworld domain. After this, a large chapter discuss several
theories of emotions from a multidisciplinary point of view. Following
this, we have a chapter on adaptive agent architectures and the
importance of the support of the concept of “value”. Trial and error
learning is discussed along with natural reinforcement learners and
artificial reinforcement learners.

One entire chapter is devoted to the notion that a society of mind
requires an economy of mind. This means that in MAS (Multi-Agent
Systems) allocation and reallocation of scarce resources must be done
using a universally accepted “value”. This value controls the flux of
resources between agents, just like money controls commodity
exchange in human economies.

The following chapter introduces reinforcement learning to the AFP
(Attention Filter Penetration) theory. Using concepts explained
previously, the author elaborates a theory of valenced perturbant states
that claim to resolve some of the problems of the interrupt theories. In
one of the appendixes a trace and some implementation details of the
MINDER1 agent is given.

Oh boy! This one wasn't easy to summarize.
This thesis presents a lot of material about an uncommon research area
in AI, that of emotions, proposed architectures and simulation in
computers. In several aspects, this is a multidisciplinary area involving
Psychology, Philosophy, Cognitive Science and others. The importance
of the subject cannot be overestimated and I finish this report
with a phrase that seems to condense this:

“Desires provide the ends, thinking provides the means”
Robert Sokolowski in Natural and Artificial Intelligence

Seminários Digitais Intelliwise

Seminários Sobre Inteligência, Criatividade, Ciência Cognitiva
Assista a esses seminários via Internet, sem sair de casa!
Clique Aqui Para Saber Mais!