CSCE475/875
Multiagent Systems
Handout
19: Collaborative Topic Summary Assignment 6: Q & A
November
13, 2009
>> Several
questions were selected from Assignment 6.
Here is my response to these questions.
Q1. What level, of an n-level agent, would be equivalent
of human intelligence? What kind of storage structure would be used to simulate
human intelligence?
Response: Human intelligence is
limitless, from one perspective. If we
say that what would be considered to exhibit human intelligence, then as easy
as 1-level agent can suffice. Remember
the ELIZA system that I talked about—the one that passed the Turing Test? That system was simply a 2-level agent,
modeling what its counterpart was saying.
There is another issue here: the higher the n, the more
complex the system is. So, we should
also consider whether there is a need for n, or n-1, or n-2, etc. Practicality has to be factored in when
designing a MAS.
In terms of storage structure, some have argued that
a hierarchical structure with IsA-links and HasA-links is sufficient to represent states of human
reasoning or represent domains of human knwoledge. Note that the question here used the word
“simulate” human intelligence. To
simulate human intelligence, I do not what storage structure would be.
Q2. Are some learning techniques preferable in
dynamic or dishonest agent environments where unlearning may become necessary?
Response: Let’s clarify something first. It seems that the assumption here is that in
dynamic or dishonest agent environments, unlearning might be needed. I see learning differently. And dynamic or dishonest agents are two
different scenarios. If there are agents
who are dishonest, if I am learning about them, then I do not have to
unlearn. Sooner or later, I will learn
that those agents’ information cannot be
trusted as I find out the utility of the actions that I have carried out based
on information is low, for example. Now,
if you are referring to information that I obtained from them, and now I need
to discard that information and thus this is
considered “unlearning”, then a fault-diagnosis-based approach can be
used. Find the faulty information
component that caused a negative utility result, for example, and then remove
that information component. This is
actually related to Truth Maintenance Systems (TMSs) in AI. Very interesting area. Now,
with “dynamic environments”, well, agents are supposed to behave in such
environments, and many learning techniques such as the Q-learning-based ones
are exactly suitable for such environments
Q3. Two similar questions: (a) In Market
Environment, is it possible to have agents with more than 2-level? (b) In
market modeling, can an agent model multiple levels of agents? Are there any
levels where the agent models others as at least the same level as itself.
Response: Let’s see this
first: A 3-level agent would be
what? A 1-level agent models 0-level
agents. A 2-level agent models 1-level
agents. A 3-level agent would have to
model 2-level agents. Would that be
useful? Think about the need for
2-level. In this case, a 3-level agent
would model 2-level agents. But I
2-level agents only model 1-level agents, should the 3-level agent even bother
to model those agents that way?
It will be difficult for, for example, a 2-level
agent to model another agent as a 2-level agent. Why?
There would be an infinite loop.
Think about paper-scissor-rock the game.
Let’s say, I model you as a 0-level player. That is, you simply choose a shape without
considering what I would pick. So, then,
I can say “You are most likely to choose paper.” And then I can follow with
“scissor”. But, if I start to model you
as a player who also models me, then I start to get into an infinite loop as I
cannot arrive at a stable choice. The
same thing applies to the market environment n-level agents. Note that in the market environment, it is
okay to have ALL agents to be 2-level agents, for example. But it will be difficult for a 2-level agent
to model another agent (even if that agent is a 2-level agent) as a 2-level
agent.
Q4. In artificial intelligence, there exists a
notion of state which includes a description of the current
environmental conditions, the known contents of that environment and the
current value of any variables that are internal with respect to a particular
agent. As the environment and other contributors of state become more
complex, the state representation becomes more complex. Furthermore, the
number of states and the complexity of relationships among states can
increase. Is there a better way to represent state? In real life,
we often only consider all that might be involved in a particular state in time
when we take care to recognize what's actually going on in our
environment. We ignore many things; for example, one state might be
considered as a span of time rather than a particular instant. We may
ignore passing cars on a highway, or apply a general concept to represent the
system that includes the highway and the cars moving along it. How might
a state be represented flexibly enough to represent functional information or
particular general models, or exact state quantities? If the knowledge
representation is different between agents, do they have to learn to understand
the semantic meaning of the messages first when they are communicating for
learning?
Response: This is a long question. In AI, state representation has always been a main
research area. The frame problem, the ramification
problem, and the representation problem are the three main problems in state representation.
Yes, in real-life, we have
to make assumptions in order to function. And yes, we do ignore many things in order to make
decisions, to act, and to reason.
To allow for reasoning with
functional information, particular general models, or exact state quantities, we
usually use a state hierarchy. Sometimes,
we use ISA links, HAS links, causal links, co-occurrence links, etc., to connect
states. We also use belief-desire-intention
(BDI) to represent our confidence/perception of states.
For the last question, yes,
if the knowledge representation is different, then either there is an ontology that
maps between the two representations, or the agents do
have to learn to do that. There has been
some work in ontology learning in multiagent systems. Quite interesting.