CogSci Summaries home |
Maes, P. (1990), Situated Agents Can Have Goals, Robotics and
Autonomous Systems, 6:49-70.
author = "Patti Maes",
title = "Situated Agents Can Have Goals",
booktitle = "Designing Autonomous Agents",
pages = "49--70",
year = "1990",
editor = "Patti Maes",
publisher = "MIT Press",
summary = "spreading activation networks. runtime arbittraton
among actions with respect to goals of system in
Author of the summary: Jim Davies, 2000, firstname.lastname@example.org
Cite this paper for:
- deliberative thinking paradigm
- emergent functionality
- Existing situated agent systems have problems with goals
- What an action selector should do [p52]
- goal-orientedness is the opposite of opportunism
- sticking to one goal as an emergent property
- deictic representation: describe only the immediate environment
so that you don't need a new symbol for every thing on earth. [p67]
- indexical-functional representations
This paper argues that situated agents suffer because they do not have
a goal structure and because the action-selection must be
pre-compiled. A novel approach is presented where "the action
selection is modeled as an emergent property of an
activation/inhibition dynamics among actions."
The big problem: What does a system do when? To solve this the
deliberative thinking paradigm was created, with a symbolic planning
system with goals and subgoals. It was found to not work well in
complex, dynamic environments due to brittleness, inflexibility and
This led to the introduction of reactive systems, situated automata,
situated agents, interactional systems, routines, subsumption
architectures, behavior-based architectures, universal plans, and
action networks, to name a bunch. [references for these architectures
are in the paper.]
These new architectures were characterized by:
Emergent functionality was an important idea shared by them.
This means that the functionality of the agent is not expressed in its
behavioral policy, but is only apparent when operating in a complex
- direct coupling of perception and action
- distributedness and decentralization
- dynamic interaction with the environment
- intrinsic mechanisms to cope with resource limits and
In the system presented, activities are not hard wired and are not
Existing systems have problems with goals:
Action selection should demonstrate: [p52]
- situation desciptions are exclusive so there is never a
conflict of action [p50]
- the rules are hand-coded (which results in domains that are too
specific) or "compiled on the basis of a
declarative description of the desired behavior of the agent."
This is no good either because the information relevent to the
choice of action is available only at runtime- the programmar
really cannot anticipate it.
- goals are crucial because: [p51]
- You shouldn't have to reprogram the agent to change the goal.
- goals cut down the space of possible actions
- complex agents have many complex goals. Goal mediation
and interaction is important.
- they are important for self-consciousness and improvement
"The hypothesis we are testing is whether action selection can be
modeled as an emergent property of an activation/inhibition dynamics
among the different actions the agent can take." [p53]
A competence has a condition-list, an add-list (to add facts to the
state of the world), a delete-list, and an activation level. A
competence module is execuatble if all the preconditions are
true. There is an activation threshold.
- favor goal satisfaction, particularly actions that satisfy more
than one goal.
- exploits opportunities, can change current goal focus when
- generally sticks to a goal rather than changing all the time
- looks ahead for future difficulties and opportunities (plans)
- gracefully degrades
- reactive and fast
Competence modules (cms) are linked:
Where activation comes from
- Successor links: links any cms from y to x where something in
the add list of x is a condition of y.
- Predecessor links: from y to x whenever there is a a successor
link from x to y. [p54]
- Conflicter link: links any cms from y to x where something in
the delete list of x is a condition of y.
Links make cms inhibit and activate each other. The observed situation
activate cms with partial matches in the condition slot. Goals are
also a source of activation. There are once-only and permanent goals
(achieved continuously). Goals activate cms that have that goal in the
add-list. Already achieved goals inhibit cms that would undo them.
Executable cms spread activation forward through successor links,
non-execuatbles spread activation to predecessor links. Conflicters
inhibit each other. There is decay as well. [p56] Activation is 1/n
where n is the number of propositions in the relevent list. This evens
out the amount of activation for cms with more or fewer preconditions
or whatever. The effect of a proposition is spread throughout the
things it's linked to.
You can change a parameter to make it more goal-oriented and less
The system shows interesting biases. It favors sticking to one goal
because when you act on one line of solution, you get closer to that
solution's completion, and the way the activation works, that line of
action will win out, generally. [p62]
The "thoughfulness" can be changed by the parameter that determines
how long activation spreads before a decision is made. Longer time
means looking farther ahead into the future. [p66] A long time makes for a
closer-to-optimal solution, but this is no good in a rapidly changing
environment. More thoughtfulness is also a speed tradeoff. [p67]
Why won't it run into the same problems as AI planners? [p67]
The system has no variables. This means that goals cannot be specified
with variables (go to x.) They get away without them because they
focus on the immediate environment, like
the-spray-can-in-my-hand. This is a deictic representation.
- algorithm is less costly. It does not keep track of a
hypothetical future state or partial plan.
- The system is not restarted at every timestep
- All interactions are local, and would be a good candidate for a
deictic representation: describe only the immediate environment
so that you don't need a new symbol for every thing on earth.
- occasionally gets into loops (because it has no memory of what
it has done.
- How to select parameters?
Summary author's notes:
- looks a lot like a production system to me.
Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
Last modified: Mon Feb 28 15:56:14 EST 2000