ARTIFICIALLY YOURS
by
CONTENTS
Sharing and access of pertinent information is critical to business functions. Now, since networks that constitute our communications infrastructures, ie:- telephone, television, computers, are increasing and becoming more automated, the notion of autonomous agents negotiating with one another in this environment is one of the most important concepts to emerge in Computer Science in the 1990s.This technology provides increasing benefits, not only to businesses and financial institutions, but also to the normal average person that utilises its potential.Can you imagine being able to access any information just by a touch of your fingertips? Or Being able to book the first available flight out of London by accessing an Airport Manager program.However, before we jump headlong onto the Artificial Intelligence bandwagon and say "OK!!! I'm all for it !!!" , there are a few questions that we need to ask. How can agent technology solve problems faced today? What are the benefits, rewards, and return on investments? What developments are in the pipeline for tomorrow? What are the opportunities and the pitfalls for software agents? Would something like Termonator 2 happen to us in the near future? Hopefully, this article would shed better light on this subject and prove that there is no stopping Artificial Intelligence from playing a bigger role in this brave new world of information consolidation.
Agents are self-contained problem solving entities which would exhibit the following properties:-
Autonomy means being able to solve tasks without direct intervention of human or other agents and also to posses a degree
of control over their own actions and internal state.
Having social ability means an agent can interact with other artificial agents and humans in order to complete their own
problem solving and to help others if they deem it appropriate.
Being responsive would mean able to react and respond to different environments.
Proactiveness is where agents exhibit opportunistic, goal-directed behaviour and take initiative when it is appropriate.
With these properties , agents reason out, plan, interact and use the wide range of information sources available,
predict outcomes and negotiate with other agents to acheive its goals.Each agent would have unique problem-solving strategies.By adjusting the rules of these public "behaviour" (called rules of the game)
, designers of intelligent agents can shape certain kinds of desirable social "behaviour" he wants the agent to adapt.
This modifications are done cooperatively, so that agents can communicate easily with one another.
Self Centered Agents ?
One of the problems that designers face is concerned with the ways which automated agents deal with one another. In domains where agents represent different world entities
such as different companies with different set of goals, self-interest prevails.Agents cannot be assumed to use cooperative negotiation strategies, but only the stategies it has been preprogrammed with and use that to get the
highest possible utility for itself without concern for other agents' utilities.Just imagine if we get a profit-interested agent
negotiating with an agent based upon centralised industry over prices of labour.One agent would negotiate high while the other negotiates low.In this simple example,we can see that these two agents
would never come to a final decision, ie:- Deadlock.In order to solve this, the industry have to find out a common ground as a basis for negotiation.
Different strategies and protocols using marginal-costs have been tried out with admirable success.Other advancement have
also been tried out but they are still in the experimental stages.
Advanced agents
Knowledge-based scheduling agents are one of the advancement being experimented with.An example of this system have been
set up by the Massachusett Institute of Technology called the Distributed ARM (Airport Resource Manager).Each agent in the ARM
domain posseses its own resources (gates,baggage,etc.) and is responsible for its own schedule, but it also can request
other agents' resources based upon cooperative agreements, ie:- Borrowing resources.By coordinating the indvidual scheduling
efforts so that each agent understands the probable requirements of other agents, the likelihood of an agent being able to find
additional resources has increased.This testbed has been succesfully distributed as a community of two or more agents.The primary difficulty faced was that no agent posseses a global view of the problem space,since only
two agents are borrowing and lending at one time.This causes backtracking, ie:- time wasting, which ultimately leads to
low-quality scheduling.
Another advancement involves self-interested agents that are able to "learn" during run-time.Here, agents are being developed
to be able to learn about the environment as well as each others strategies.Reinforcemnt Learning (RL) is based upon the idea that
when an action produces a favourable result,it should be strengthened and weakened when it does not.A learning algorithm
with this idea is called Q-learning, and these algorithms can be used on-line,ie:-repeated games against unknown opponents.
Agents of these sort have been tried out against self-interested opponents as well as Q-learners.In the former case, the Q-learners
fared optimally but in the latter,it proved to be more difficult to come to a decision.Imagine pitting Hitler against Hitler.We would probably get a double Blitzkreig.
More research need to be done before optimal use of Q-learners can be used in the real world.
Intelligent agents in the real world
Since the more sophisticated agents are still in its experimental stages, we will only look at the earlier models of working
intellgent systems.One example would be the Contract Net Protocol (CNP) [Reid G. Smith,1980] which introduced a from of simple
negotiation among cooperative agents, with one agent announcing the availability of tasks and awarding them to other business agents.
By introducing a more sophisticated economic model of NCP,Thomas W. Malone, refined this negotiating technique, proving
optimality under certain conditions.Malones research introduced a motivational framework where agents negotiate on the basis of
an economic model.
Another step in dealing with the automated negotiation domain is the work on TRACONET [Thomas Sandholm,1993].In this working model,
self-interested agents make announcing,bidding and awarding decisions using a simple statistic approximation scheme for
marginal costs calculations.Nevertheless,in the working context,agents are still assumed to have perfect rationality and not
assumed to exhibit uncertainties.TRACONET was designed so that commitment,(when one agent binds itself to a potential contract
while still waiting for a reply) took place in the bidding phase.In the real world, all decisions should be made in real time, so
this was a giant step of acheiveing this goal.
CONCLUSION
There are plenty of technological advancements at present in leading to the ultimate goal of having intelligent agents making optimal
decisions for us.On this basis we can safely say that businesses can benefit from this increasing speed of interaction and can take advantage
of this increasign returns to scale.As we have seen, most intelligent agents need a basis to negotiate with, this being an economic type.
So,in the near future, hopefully we can see fully working learning intelligent agents which can predict and decide tactfully
and accurately for businesses to reap its goals.This would surely make a lot of poeple unemployed.We would not want that to
happen now, do we ?????
References
Distributed Artificial Intelligence Laboratory, Massachusette Institute of Technology.
Jeffrey S. Rosenschein and Gilad Zlotkin,R. 1993. Designing Conventions for Automated Negotiation.Abstract from AI magazine, Fall 1994 pg. 29-45
Jeffrey S. Rosenschein and Gilad Zlotkin,Rules of Encounter, MIT Press,Cambridge,Massachusetts,1994
Distributed Artificial Intelligence Laboratory,Hebrew University, JCerusalem,Israel.
What is an agent?