Some of my views on progress in AI and what could be in store are in this article published on 18th May 2012...
Tate, A. (2012), Artificial Intelligence - from Fantasy to Reality, OmegaScience, On-line Magazine. [Local Copy]
Artificial Intelligence is the science and engineering involved in creating systems which perform in ways which would be considered intelligent if performed by people.
~What is your particular speciality in Artificial Intelligence?
AI Planning and Activity Management. Some of the work is described at http://www.aiai.ed.ac.uk/project/plan/
~Why did you choose a career in AI
research / development?
I wanted to be a scientist from my Junior school days (1959), and got
interested in computers during my secondary school years (1966-8).
When going to University, I wanted to work on the software aspects of
computing and chose one of the two UK University courses at the time
that did this (Lancaster, the other possible was Essex). My teachers
there started a new course in Artificial Intelligence and had contacts
with the Edinburgh AI Department. I started work on an undergraduate
project on search methods and planning in 1971. This involved me in
talking with Prof. Donald Michie in Edinburgh who was working in that
area. He had been a war time code breaker at Bletchley Park working
with Alan Turing. He invited me to join him in Edinburgh as a Ph.D student and
things carried on from there.
~What is the AI system you are researching/developing designed for?
To support command, planning and plan execution within organisations such
as:
~What is the most exciting part of AI that
encourages you stay in the field?
The field is only just seriously beginning in my view - and it's poised
for a massive development. There are very many more people involved in
AI today than was the case even 15 years ago. The successes of AI
to date (and only some are widely recognised) have formed a solid
basis for realistic exploitation and excellent prospects for future
development.
You will see deep space probes with advanced automation and AI travel
out from our planet sending back home exciting discoveries, you will
see autonomous sea, land and airborne vehicles
exploring parts of our own planet too inhospitable for man to travel
there. You will see humans and robots working alongside one another
in emergency and rescue situations.
You will be able to have a personal assistant or co-worker who
will work alongside you, get to know your tasks, processes and
preferences. It will do those things you wish you had time to do
yourself but which are never at the top of your agenda. The same system
will adapt itself to becoming an active aid as you and your family age.
Someday, it might even be able to draft an answer to an e-mail message
like this one, as it will know the person well enough.
~What subjects would you encourage
high-school level students to take, who are interested in AI?
Computing is only a part of AI. Look to subjects that encourage problem
solving, modelling and understanding other people's skills and their
roles. Computing, Maths (especially Logic), English (especially
Grammar), Psychology are topics that I would suggest. Only some of
which I studied at school and in my first degree at University. I much
regret not having a more solid basis in language grammar for example.
~Other Comments:
I wish you well with your studies and this survey. I remember very well
my own school projects on stroboscopic photography of gymnasts, on
the industries of West Yorkshire (where I come from) and on decay in
radioactive materials. I keep some of the records of those experiments
and have fond memories of the teachers who supported me in my early
school days. I hope you have as much fun as I did.
~What are the negative aspects of
artificial intelligence?
It is in the nature of any technology that it can be used for good or
for bad. If AI systems are seen as agents of their creators or the
groups that deploy them, then accountability must reside with those
that put such systems into use. That is a general concern we all should have
whether we are talking about tools such as a hammer, a car, or a
computer system such as an elevator controller, an automated financial
stock dealing system or a computer controlled gun.
The rather special concern that I have about computer systems generally, and AI systems as one case of those, is that people may choose to deploy such systems and give them autonomy or authority to perform believing that they understand the range of behaviours possible and programmed into such a device. As with letting loose any other agent (whether chemical, biological or human) we may not be able to predict the outcome of such a delegation of authority or handover of control. Due to the technological complexity of such systems, and the commercial pressures to push solutions on clients such inadequacies or problems may not be understood by those choosing to deploy them.
~In the articles and the literature that we have studied, it seems that the interest for, and development of, expert systems have decreased during the 1990s. What is your opinion about that?
The term expert systems is narrowly applied to rule-based systems incorporating human knowledge in a form that can be used for tasks such as interpretation and analysis. The more general term of knowledge-based systems is applied to the broader use of knowledge in systems that utilise a variety of reasoning mechanisms.
Academic interest in rule-based systems is lower than 10 years ago, but applications continue to grow. KBS work and knowledge management work has really just begun and will explode. It will in my view become a dominant technology for any organisation in future.
~When did AI start to become a popular subject?
~How much farther do we have to go?
Artificial Intelligence has been popular since the very start of the
computing age in the early 1950s. Early AI work drove some of the
early languages developed for general computing, and continues to do
so (e.g., in object-oriented languages and Java developments)
Personally I think we are just seeing the real start of the interesting developments in computing, and AI has barely begun in terms of where we have to go.
~What have been the latest successes in AI?
~How realistic has AI gotten?
~Are the AI programs really intelligent?
~Could the computer take over the human race?
~How could AI affect the future of mankind?
Well, in short: many and varied; not much at all (yet); no (not yet);
depends on us; in a profound way.
~Is artificial intelligence the computer
itself or is it the concept of trying to get computers to "think"?
It is the study of getting computers (artifacts) to act intelligently.
~What would you say is the most
"intelligent" computer yet made? Why is that computer considered to
be intelligent?
The intelligence is normally in the software which runs on the
computer. There are many very clever, knowledgeable or intelligent
programs working on well defined problem domains and helping people
in their everyday lives without them being aware of it. Even the
fuzzy logic controller found in some camcorders and some elevator
control programs that get the elevator to your floor more quickly
originated from AI studies. But powerful brute force systems like the
core of IBM's Deep Blue (the chess playing program/computer) do not
strike me as very intelligent. They are just powerful and able to
search many possibilities in a short space of time. Deep Blue though
did have some intelligent aspects bolted to this core brute force
approach. I would describe as more intelligent the software on board
the Deep Space One spacecraft which has some very fine intelligent
planning and control software on it as it continues it's mission.
This AI system was created by NASA's AI team at Ames and Jet Propulsion
Lab and controlled the spacecraft autonomously for 2 days in 1999 as an
experiment.
~Does our cultural fixation on the
"anthropomorphic view of A.I." seems ridiculous to scientists and
theorists in the field.
I personally am a fan of science fiction too, and I have no problem
with the anthropomorphic viewpoint for fictional purposes.
~Is the anthropomorphic aspect of A.I.'s
development one that excites a specific sect of the scientific community? Do
you happen to know whether the "intelligence" part of the equation is ahead
of or behind the delivery system, ie., the robotic body? Are we likely to
have high-functioning computers with A.I. long before we start creating
little boys?
Yes. Physical and reasoning aspects both require much
work. Interacting with the world in a useful way is a big
challenge. We are a long way from realistic biped bodies that move in
a humanistic way - the closest are some biped robots at Honda in
Japan. Roving robots (as opposed to the more typical fixed position
robot arms commonly used in industry) are limited to simple tin can
shapes with wheels or rovers with tracks today. A lump of metal on
rubber skids is not quite up to appearing as a guest artist in a film
though - well I suppose R2-D2 did exactly that;-) Robots today are
more along the lines of the Pathfinder rover that landed on Mars.
That is a more realistic embodiment of the ambulatory and sensing
capabilities a year 2001 robot might expect to be able to use.
Disembodied AI and knowledge-based systems with high levels of skill and knowledge already appear in deployed applications in everyday use. Some examples of work from my own Artificial Intelligence Applications Institute (AIAI) are listed at http://www.aiai.ed.ac.uk/i3ai/. There is a long way to go to address all the very many varied aspects of human intelligence.
~In 2014... Tesla and SpaceX CEO Elon Musk
expresses concern that AI is "potentially more dangerous than nukes" and
physicist Stephen Hawking warns "AI could be a big danger in the
not-too-distant future"... are we all doomed...
Oren Etzioni, head of The Allen Institute for Artificial Intelligence
(AI2), wrote a piece on
9-Dec-2014 in response to concerns expressed by public figures like
Elon Musk and Stephen Hawking about the threat that future AI systems
might pose to humanity if they decided on their own objectives and
actions. This had been picked up by the popular press at the
time. Oren sought to separate the notions of "intelligence" and
"autonomy" and explained that autonomy by even the dumbest algorithms
or systems was the threat, not intelligent systems themselves. [The
article is available
here.]
I would add that a simple "if x then y" rule can be a threat if deployed inappropriately... "if <incoming object> then <launch all missiles>". AI scientists in the 1980s had been at the forefront of warning the public and governments about the risks posed by the proposed US "Star Wars" defence system in that over reliance on the assumed capabilities of the control systems that might be utilised could pose the biggest threat. Soon afterwards there were descriptions of "near misses" in human-controlled defence systems where mistakes were made in interpreting incoming missile data, due to training simulations being left active.
~What is your definition of AI, how is it linked
to what we define as intelligence, what have been the key break throughs in
AI?:
AI is the science and engineering of getting machines to do things that a human would regard as intelligent. There are two aspects:
Human intelligence might be seen as having analytical, creative, practical and adaptive aspects. Tasks like interpretation, prediction, diagnosis, design, planning, learning. AI has made advances in all these areas.
Some AI systems may be able to perform in ways that are not human-like at all. They may use alternative forms of reasoning more suited to their own strengths and avoiding their own weaknesses.
There is a relationship to Turing Test which involves an observer asking questions of another entity without seeing that entity and deciding if it is machine or human.
There have been many advances in AI over the years... Speech understanding, industrial robot visual inspection systems, planning, scheduling, financial expert systems, medical advisory systems and so on. AI is in fact in use all around you:
There are also some general AI methods that are frequently used in many systems: for example heuristic search, constraint solving, rule-based reasoning, and adaptive algorithms such as genetic algorithms and swarm intelligence.
Areas of application:
Software agents are a way to modularise or package computer services. They can internally maintain state information and use messages to interact with the outside world. They perform their services in response to these messages. This allows for simple or more intelligent agents, and agents which can be "mobile" in the sense that they can be shifted about to execute in different locations on different platforms.
Agents/web services packaging is making AI/knowledge systems more accessible to the broader enterprise computing area - rather than the previous island of AI automation in industry and business.
~What is Swarm Intelligence
(SI):
Biologically inspired methods are used in several areas of AI - neural
networks, genetic algorithms, ant colony optimisation and swarm intelligence.
Swarm Intelligence is the emerging collective behaviour of many simple
agents. SI provides a basis with which it is possible to explore collective
(or distributed) problem solving without centralized control or the provision
of a global model. SI is suited to optimisation and other adaptive tasks.
~Can AI have emotions:
It depends on whether you see emotions as something only humans can
express... if they are fundamental aspects like desire, fear, pleasure
and pain these can be used as guides to influence programs and robots
to perform in real environments which can be sensed by the robot.
Computers that are aware of their users' emotions could improve their dialogues... and it could be a useful mode of expression to engage their users more effectively. E.g., Sony's AIBO dog robots include a range of facial expressions. Sony expects entertainment robots to be a multi-billion pound market in the next decade.
An EIFF 2011 event involved a screening of the 1984 scifi classic The Terminator. The event was followed by a discussion with two real-life AI researchers from the University of Edinburgh School of Informatics. We managed to catch the two scientists before the event and talk to them about the past and future of AI, its perception in the media, and whether we should worry about the Robot Apocalypse. Listen to our conversation with Austin Tate, Professor of Knowledge-Based Systems, and Sethu Vijayakumar, Professor of Robotics.~Where can I see more about AI and Robots in Movies?[EUSci Podcast MP3] [Local Copy: Image of Web Page, MP3]
Interviewed on 24-Aug-2012 for Compucast monthly podcast.
1- What is AI planning? It is a branch of artificial intelligence but can you please express what aspect of human behaviour is captured by the AI planning and how the task is formalized?
Planning is one of the most important aspects of intelligent behaviour. The ability to identify and select appropriate activity and to project forward the consequences of executing that activity is fundamental to humans and intelligent computer systems alike.
Essentially planning is the process of identifying some objective or goal and using a set of activities or capabilities which may transform a known situation into the desired situation.
2- How has the AI planning evolved? do we have a classical Vs. modern division in terms of the techniques which have been developed over the last couple of decades?
Classical planning is a term used to describe some simplifying assumptions to mqake planning problems more approachable in computation terms.
Assumptions can include instantaneous action effects, unchanging state except when actions are performed, and even lack of concurrent activity.
An alternative is to use a much richer representation of the world and the constraints in that world, using time, resource and spatial constraints which can radically restrict the solutions possible. This is an area called knowledge-based planning. In some cases, with good constraint reasoning, search can be reduced to a small number of legitimate solutions. In some domains a mixture of human and machine problem solving can work effectively together to limit the choices explored.
3- what is the computation and space complexity of solving planning problems? do you also deal with the problem of state space explosion or similar problems?
Even very simple and basic planning problems can be computationally very difficult, and can lead to very large or infinite search spaces. If a problem is stated in a mathematical puzzle like way apparently trivial block stacking problems can be beyond solution by classical planning methods.
Early AI planning techniques typically had a search space where nodes were application domain state snapshots, and the transformations were the activities in the application domain. This is called "state-space search". Searching this state-space can still be useful in some limited cases and where very efficient modern algorithms can readily seek solutions in such a space.
But many more recent planners search quite a different space. Nodes are partial plans, and the transformation are planning related processes, such as expanding an activity to a lower level of detail, or adding an activity to achieve an unsatisfied goal. This is called "plan-space search". The planner reasons about the plan as a whole rather than specific application states. It can basically debug an approach to how the plan will work. Reasoners and constraint solvers can operate within this framework to establish what facts hold where needed, or to prune search where things are not possible.
4- some recent applications: from simple applications to space missions. How is AI planning used? do you have a concrete example?
AI planning has been used on a very wide range of applications, and is in productive everyday use. Some we have been involved in with Edinburgh AI planners included:
Other practical planners from elsewhere have been used in things like
5- How can, for instance a robot, make useful plans when all its surrounding might change by the time it has taken the plans into action? How does AI planning deal with the uncertainty of the environment or the fact that robot's environment is constantly changing?
Many practical planning systems have to handle real execution in real environments, so changes of the environment, and even changes of objective have to be handled. If the problem is seen as a continuous task of adopting and adapting a plan or behaviour to the circumstances, rather than a classical single shot goal achievement problem, you can begin to address this dynamic planning problem.
For some time our planners have been able to monitor the world they are operating in so as to recognize when necessary conditions and constraints are not met or have been violated. This allows for plan repair or complete replanning to occur often far ahead of the problem becoming an issue.
6- Some recent projects that you have been working on recently? (the initiative to improve public awareness about planning?)
My recent work has been on supporting distributed communities in responding to emergencies using a mix of AI planning, social networks and virtual collaboration environments. We are about to offer a massive open on-line course - a MOOC - in AI planning to make the techniques be more widely known and available.
~What are the best graduate / other schools for learning AI?
Some of the world leading centres for AI are MIT, Stanford, CMU and
Edinburgh, though there are some fine smaller departments and
individuals researchers and teachers all across the globe. Nowadays
it is also possible to study via free open online courses (MOOCs)
such as Udacity's
Introduction to
AI
~What are some of the top games using AI?
Take a look at this video and resource page from Alex Champandard
created for our AI Planning MOOC:
http://aigamedev.com/open/review/planning-in-games/.
