Sixty years since Alan Turing asked ‘can machines think?’ artificial intelligence has taken giant leaps. There are some ‘if’ factors though, which pose fundamental problems in the way for intelligent geospatial systems.
HAL, the artificial intelligence that controls the spacecraft and interacts with the astronauts in director Stanley Kubrick”s film 2001: A Space Odyssey, was capable of remarkable introspection: “I know I”ve made some very poor decisions recently, but I can give you my complete assurance that my work will be back to normal. I”ve still got the greatest enthusiasm and confidence in the mission. And I want to help you.” That was way back in 1968.
Earlier in 1950, when mathematician and computer scientist Alan Turing asked “can machines think”, he envisioned a machine that, by its communication skills, cannot be distinguished from a person. (Fig. 1.)
So, how far have we got with intelligent systems, beyond some marketing hype? Is there any agreement on some measurable criteria to what extent, or whether at all, a system can be called intelligent? And what, in particular, is an intelligent geospatial system?
HAL is a wonderful example of humanoid communication skills, including learning mechanisms as reflected in the ability for introspection above. We are nearing 2012 end now, and yet the best we can say on this is artificial intelligence has had a rather chequered history of successes up to now. It works in rather narrow areas, such that we “now have machines that have trumped human performance in such domains as chess, trivia games, flying, driving, financial trading, face, speech, handwriting recognition…” Some of these narrow areas either are within the geospatial domain, such as flying or driving (positioning, motion planning), or have applications in the geospatial domain, such as speech recognition (humancomputer interaction about places, times, or directions).
HAL-like navigation system
To review the current abilities of geospatial systems, let us imagine a HAL-like navigation system, HAL-NAV. Its role would be solely to give advice to the traveller; thus, it would be able to have any conversation in this context. Such an idea is very similar to the restricted Turing test, an award advertised annually by Loebner “to find the world”s best conversational computer programme”. For any chance of success, let us keep HAL-NAV narrowly focused: the communication between a driver and a car, or a passenger and a tram would not be about life or love, but restricted to wayfinding, and perhaps the weather, traffic, meetings, or the shopping list.
Research in human cognitive capabilities and linguistic behaviour inform the design of formal models to reflect these capabilities and behaviours in a computer. However, the Turing test only relies on behaviour; it does not require a computer to internally function like a human being and accordingly, the formal models do not aim to explain the functioning of the mind.
Nevertheless, a whole range of challenges have to be solved in order to build HAL-NAV, among them understanding verbal or graphical human place descriptions, context and relevance, embodied experience of environments, especially salience, modelling the notions of place and modelling user-centred time geography and personal preferences or knowledge of an environment.
Let us focus on a few of these challenges. Consider a request a traveller might pose to HAL-NAV: “Can you tell me the way to the airport?” Surely, people will not have difficulties to answer this question — it is easily understood and applied by the traveller.
In some respects, it is not too difficult a question for HAL-NAV either. HAL-NAV will be superior to humans in computing routes: compute faster, process more data and produce more accurate results. HAL-NAV can guarantee to compute the quickest route and perhaps even include data about current or past traffic in this computation. It can also compute the cheapest route, and the simplest route. Now, here is the first challenge for HALNAV: which one of these does the user want? The answer leads to a territory still not well understood in artificial intelligence research: context dependency of optimal states. The regular commuter may simply want an update based on the current traffic situation, the business traveller wants to be guided to the simplest route to the rental car return at the airport and the backpack tourist wants to know how to reach the airport via the cheapest way such as public transport or by ride sharing. Translating these issues to current commercial navigation systems, challenges take the form of dealing with user preferences, integrating multiple modes, thus accessing distributed, sometimes decentralised data sources and in doing so, maintaining privacy.
Next to the challenge of understanding and dealing with context, HAL-NAV has to correctly interpret the destination description “to the airport”. But the way people describe their places or their travel destinations are infinitely rich. So, while HAL-NAV would have the transport network data readily available, it may still feel challenged by place references. Gazetteered place names, business directories and points of interest are a starting point, but this data does not cover the vernacular (non-gazetteered) names, vanity addressing, extent of places and indeterminacy of some places, ambiguity of place names, and the interpretation of qualitative relationships between places.
In addition to the challenges of understanding the users” request, HAL-NAV is also challenged by responding to the user in a way that is easily understood and easy to follow. For example, a person may respond to the above question like: “Okay, go down this street [pointing], at the traffic lights down there turn right, then drive six, seven blocks until you reach City Gate, a building characteristic for its shape, where you turn left on the highway. Follow the highway to the airport.”
How is HAL-NAV different?
Why is this description a good one for the traveller and tough for HAL-NAV to replicate?
For the traveller, the description conforms to Grice”s conversational maxims, especially to the maxims of relevance and of quantity (“make your contribution as informative as required,” and “don”t make your contribution more informative than is required”). It also applies some other remarkable principles. For example, it mixes conversational modes between pointing and spoken word and relates to vista space, engaging with the user in an embodied manner. It also lowers the focus on numerals by deliberate use of uncertainty — still giving a rough idea of distance, but not stretching numerical cognition and short-term memory — and resolves uncertainty by a landmark instead (thus adding even some redundancy for the comfort of the user). Finally, it chunks all further instructions along the highway into one, relying on “knowledge in the world” — the signage to the airport — thus avoiding redundancy in the triangle between the traveller, the environment and the speaker.
The problem for HAL-NAV is machines lack understanding for landmarks. From a cognitive perspective, landmarks are the elements structuring mental spatial representations. They are closely related to the embodied experience of environments. Some of them are personal (“the place where we met”), others are shared (“the red building”). To be a landmark, a geographic feature must stand out from its neighbourhood by some sensible properties, and this will always depend on the focus of the person. These observations about landmarks mean one can only formulate relative classification schemes for landmarks, and these schemes will need to explore rich data resources and act with context awareness.
The simple question of a wayfinder has already brought up a few interesting challenges for intelligent geospatial systems just in the area of navigation services. There are more challenges and more geospatial systems, of course. Perhaps it is time to suggest a ”grand challenge” for the research community: a restricted Turing test similar to the Loebner Prize, here on intelligent navigation services. Has a HAL-NAV become possible?