7

"Model checking for Artificial Intelligence"

Michael.Wooldridge is Professor of Computer Science in the Department of Computer Science at the University of Liverpool. He is currently Head of Department, and in addition head the Agent Applications, Research, and Technology group (Agent ART), which carries out both pure and applied research in the area of autonomous agents and multiagent systems.
 

Robert Fisher (Edinburgh University, Trinity College Dublin)

"AI and Cinema - Does artificial insanity rule?"
"Stabler augmented reality through projective geometry"

Ronan Reilly (University College, Dublin)

"Eye movements in reading: models, problems, and prospects"

Robots and other artificially intelligent mechanisms are now a common plot device or even main character in movies [2], and have been so since the early days of cinema. A www-based survey of these movies [1] has over 80 entries, not including TV programs, cartoons, cyborgs (augmented humans - 35+), androids (synthetic humans with some unspecified computational mechanism - 20+) and still unclassified movies (115+).

Of the "true AI" movies, all but 10-20 use a "mindless" sort of AI without self-reflection or self-awareness, largely as an efficient robotic killing machine.  Although this is an exciting plot device and a worrisome possible future, these AI agents do not have much depth and ultimately will have the same interest to viewers as a well-engineered automobile.

But what of the other movies? Examples of these are Metropolis (1926), The Wizard of Oz (1938), 2001: A Space Odyssey (1968), Demon Seed (1977), Making Mr. Right (1987), Virtuosity (1997) and AI Artificial Intelligence (2001).  These movies have AI-based agents as central characters, interacting with humans in many of the same ways as other humans. The movies are fascinating, but not solely because of the AI technology and story line. Irrespective of the success of the movie as a creative work, or the plausibility of the AI agent, these movies explore interesting questions about humans and what defines humanity.

From this perspective, there is a notable fact about almost all of the AI agents in these movies: in varying degrees, all of the agents show abnormal behaviour, from obsessive to pathologically insane. The main forms of deviant behaviour are:

? Fixation on being loved: AI Artificial Intelligence, Making Mr Right, Electric Dreams
? Fixation about becoming human or at least physical: Demon Seed, Virtuosity, Wizard of Oz, Making Mr Right, DARYL, Iron Giant, Star Trek: The Motion Picture, Generations
? Extreme behaviour from irreconcilable conflicts: real life and complex society demands many different responses and actions from humans. Sometimes these are hard or impossible to simultaneously satisfy. HAL in 2001 seems to have suffered from this sort of conflict.
? Monomania: Dark Star
? Megalomania: Virtuosity, Colossus - The Forbin Project

Aren't these all common human behaviour disorders, when taken to extremes? And, aren't these also issues that we worry about when applied to humans? And properties we commonly ascribe to "other people" in prejudices that we no longer (should) apply based on nationality, race, religion, gender, social class, etc? Thus, these AI agents allow us to explore many of the same questions explored by other mainstream movies, perhaps in a context that allows us to ignore questions of racism, sexism, and so on.

While I am not deeply considering the question of whether real artificial Intelligences will always go insane, especially since the sorts of agents depicted in these movies are many years away, it is interesting to speculate about what characteristics of AI might lead to insanity on the part of the agent. Some likely types of insanity are:

? Paranoid self-preservation: because humans could 'turn off the power', they are a threat, needing control, elimination, displacement, and so on.   This is a natural individual survival instinct enlarged beyond social control.
? Flawed or ungrounded reasoning: somewhat like Hamlet, an agent following a chain of reasoning that is divorced from real data and feedback can come to odd conclusions. Alternatively, reasoning about material objects based on physical properties is likely to be stable, but what about reasoning about humans and behaviour? Even as humans, we constantly misinterpret situations because we use incorrect assumptions or heuristics. AI agents will be no different, except some of their reasoning mechanisms may be beyond improvement, for example based solely on logic.
? Flawed perception or hallucinations: sensory data, particularly visual data, is complex, confounded with shape, position, illumination and sensor range and response. Humans resolve the confusion by using a combination of active perception, knowledge of the normal world and knowledge of specific objects. AI agents using unsound processes could make seriously incorrect conclusions about the external environment.
? Superiority complex: because of the likely greater computational speeds, broader perceptual ranges and mechanical strengths, AI agents might ultimately conclude that they are sufficiently superior that they need not treat humans (nor even other AI agents) properly.

These abnormalities also commonly afflict humans in varying degrees. They arise because of deficient or defective perceptual, memory or reasoning mechanisms, or inadequate or inappropriate socialisation. AI agents are unlikely to be different. Because of their different reasoning mechanisms, they are likely to manifest unaccustomed aberrant behaviour as well, but the forms of this are harder to predict.

Coming back to the cinema, these forms of abnormal behaviour have had some but not a lot of exploration from the perspective of central characters in movies (for example, Skynet in The Terminator is the cause of but does not appear in the action of the movie). So, we might see some movies pursuing these aberrations. We might also see some movies that investigate how human society might control, repair, detect or prevent 'insane' AI agents. This could include such techniques as
? isolation, destruction and rehabilitation,
? creating an acceptance of and a way to act in irreconcilable conflict situations,
? creating an acceptance of 'death',
? engaging AI agents in a 'society' that provides social feedback, behaviour modification and constraints, or'drug' control (e.g. by producing modifiers of reasoning and perceptual abilities).

Is insanity what we can expect from any future movie having an intelligent, interactive AI-based agent? There are a few recent movies with AI agents that are not obviously aberrant or insane. One example is the R2D2/C3PO team in the Star Wars series, but these AI agents do not aspire to greatness and equality, and we are disarmed by their comic cuteness or incompetences. A more interesting example is 2010: Odyssey Two, where the resurrected HAL agrees to a heroic self-sacrifice once it is fully informed of the situation and reasons. Thus, it likely that we will also see AI agents in other situations where humans typically appear, in part to explore our own human concerns, anxieties, conflicts and histories.

On a more speculative note - why should we even assume or expect that real AI agents will behave in any deeply intelligible manner? Their sensory, memory, reasoning and physiology systems will be quite different from ours, and all these shape their intellect, much in the same way that 'identical' twins diverge throughout life and human societies develop different and often hard to understand cultures. This seems like a fruitful theme for movie exploration.

Given the majority of movies containing an unfavourable representation of AI agents, there might be problems for research labelled explicitly as AI. If the horrors of nuclear war were well known in advance of the development of nuclear weapons, there would have been stronger social controls. I am not equating AI agents with nuclear weapons, but I could imagine that a sensitized general public might feel this way. Some AI scientists also strongly promote this nightmare. Raising these issues is correct and will help us form a consensus about the allowable roles for AI agents in society. But, considering the 50-100 years likely before real agents appear that remotely approach the abilities of the agents in the movies, the movies and the concerns aroused may do much harm to AI research, even when that research is really just focussed on tools that enhance human abilities.

References
[1] http://www.dai.ed.ac.uk/homes/rbf/AImovies.htm
[2] The University of Illinois's Cybercinema page: http://www.english.uiuc.edu/cybercinema

Biography
Robert B. Fisher received a B.S. with honors (Mathematics) from California Institute of Technology (1974) and a M.S. (Computer Science) from Stanford University (1978). He received his PhD from University of Edinburgh (1987), investigating computer vision in the Department of Artificial Intelligence. Dr. Fisher is a Reader in the Division of Informatics at the University of Edinburgh and is a member of the Institute of Perception, Action and Behaviour. His research covers topics in high level and 3D computer vision. He directs research projects investigating three dimensional model-based vision and automatic model acquisition of industrial objects and buildings. He teaches general and industrial vision courses for undergraduate, MSc and PhD level students.
 
 

The normal technique for augmented reality transfer involves estimating the 3D position of the overlay imagery and then applying the standard pinhole camera model to project it into the target image. This is effective, but, because of the loss of 3D information in the target image, estimating the 3D position of the overlay can be a little unstable. This can produce noticeable frame-rate jitter of the overlay. As an alternative, we explored direct graphical transfer of planar overlay imagery, using a planar homography from the overlay imagery to the target video sequence. This reduced the mapping jitter by a factor of 3-5, leaving the error at about the threshold of perceptibility. We also developed a technique for transfer of rigid and non-rigid 3D shapes consisting of multiple linked planes, as a first step to direct full 3D overlay transfer.

The single overlay plane transfer approach uses a planar homography to map the overlay image point-by-point into the target image, using a homogeneous co-ordinate representation of the points and a standard planar homography mapping from projective geometry. The mapping is estimated using the Direct Linear Transform (DLT) method, which is described clearly in Hartley and Zisserman's recent book [2]. This algorithm requires a set of corresponding points between the source and target image planes. Four points are the minimum, but additional points are better to provide stability to noise. Many uses of the DLT algorithm use corner points, but these are hard to find in this application, so we developed an extension of the Iterated Closest Point (ICP) algorithm. This allows us to obtain a one-to-one matching between boundary points of the source and target regions. As well as increasing stability, this allows us to transfer regions with irregular and curved boundaries. The normal ICP algorithm matches rigid shapes using Euclidean transforms (rotation and translation); our extension substitutes a full projective transformation. The projective ICP algorithm works effectively provided a reasonable initial estimate of the transformation is available (which is also the case with the standard ICP algorithm).

The extension to partial 3D transfer uses independent homographies for each plane, which are then constrained to map common points in the overlay image to the same point in the target video. This reduces the accuracy of the transfer of any individual plane, but ensures that the planes do not drift with respect to each other. There is a slight increase in mapping jitter, but the constraints are satisfied.

These new techniques enhance the repertoire of methods for producing high-detail, stable augmentation of video. More details can be found in [1].

References
[1]  R. B. Fisher, "Projective ICP and Stabilizing Architectural Augmented Reality Overlays", Proc. Int. Symp. on Virtual and Augmented Architecture (VAA01), Dublin, Ireland, pp 69-80, June 2001.

[2]  R. Hartley, A. Zisserman. Multiple view geometry in computer vision. Cambridge; New York: Cambridge University Press, 2000.
 

Eye movements in reading: models, problems, and prospects

Ronan Reilly
University College, Dublin.

Back to the AICS 2001 page.