1. Reasoning about the external environment should be minimal. For example problems involving representation and inferences about moving 3-D shapes should be avoided. This rules out control of manipulators e.g. performing assembly tasks.
Motion on a 2-D surface of objects regarded as shapeless points should provide a rich enough world for initial tests.
2. Simulation of a number of intelligent autonomous agents should be avoided. The primary goal is to understand how a single intelligent system can control (or be composed of) a number of independently active parts. This task will be difficult enough without having to cope with simulation of several different intelligent behaving systems. I think this rules out traffic control problems and a number of other interesting problems.
So all the agents should either be fully under the control of a single intelligent monitor or else should be very simple, and deviations from predictability should be random, to minimise the programming required.
3. The behaviour of both "internal" and "external" objects should be capable of giving rise to new conflicting goals of varying degrees of urgency and importance at unpredictable times.
4. The planning and decision making processes should not be trivial.
5. From the point of view of Sussex research interests the design should illuminate some of the cognitive functioning of an intelligent organism with a range of goal generators, preferences, policies, etc., in a rapidly changing world with opportunities and dangers. Theoretical work already in progress suggests a range of design constraints and partial solutions. (E.g. Paper by A.Sloman at BCS Expert Systems 1985 conference.)
Babies arrive in batches from time to time, have to be looked after until they have reached a certain age, and then are sent elsewhere.
The world is a field divided into rectangular play areas separated by ditches and joined by bridges. There are also some walls that prevent the whole area being simultaneously visible from every point.
There is a single nursemaid controlling a collection of nanny-robots that can move around the field, detect the location of the babies, and detect various dangerous situations as described below.
The babies and the nannies can't jump over the ditches but the nannies can see over them. If babies get close to ditches they sometimes fall in and can't be recovered. So they have to be kept away from the edges of the play areas. Babies don't detect ditches till they are very close, and then they send frequent and intense signals to the nursemaid, indicating their approximate location. If they fall in the signals stop.
Snatchers and ogres occasionally turn up and try to run off with babies. They can enter the field by jumping over surrounding ditches, but in order to get away with a baby have to cross a bridge. A baby sends a frequent and intense signal while being carried off by a snatcher or an ogre.
Snatchers and ogres are not very intelligent and always go towards the nearest baby in sight. Having picked up a baby they try to get away over the nearest of the four external bounding ditches, and always select the nearest bridge that leads towards the bounding ditch, or if there isn't one choose one at random.
Once the baby has been carried over the bounding ditch, it is lost. Hot pursuit is not allowed, for some reason not relevant now. It is not possible to destroy the bridges providing the escape route for the predators.
The nannies can move faster than the snatchers and can disable them if they get close enough. One nanny can disable a snatcher but two are required for an Ogre. If only one nanny gets close to an ogre, all it can do is prevent the ogre from moving until help comes. However, if help does not come within a certain time the nanny will be killed by the ogre's struggles to escape. While holding an ogre the nanny will send increasingly frequent and increasingly intense "mayday" signals to the nursemaid.
There are, however, a few super-nannies who can manage alone to disable an ogre. They also run faster.
The nannies have very little intelligence. They can transmit information to the nursemaid and receive information by radio. They can request plans, which they then execute. They have a limited repertoire of primitive actions out of which plans can be formulated. They also have a limited repertoire of perceptual abilities. They can report their location and the location of other objects within a certain radius, provided a wall does not block the line of sight.
Nannies can be interrupted while executing plans, and can be given instructions to abort, suspend, or modify a plan (e.g. move more quickly, change the route, apprehend a snatcher on the way, etc.)
Nannies cannot themselves modify plans if they get into difficulties. They have to get new plans from the nursemaid, after reporting the problem.
The nursemaid has a limited parallel-processing ability so that it can receive information and requests concurrently with planning and interpreting new information. Planning or reasoning can be interrupted by incoming signals, but there is a resource-limited filter for interrupts, whose filtering strategy can be adjusted in the light of the current task.
Nannies can detect babies moving towards ditches and approaching snatchers and ogres and can answer questions from the nursemaid about what they see. They can start sending information spontaneously to the nursemaid too, if a baby approaches a ditch or snatcher or ogre approaches a baby. The messages will be sent intermittently with a frequency and intensity that depends on the closeness of the approach.
Nannies use batteries that can run down, and there are a few re-charging points in the field. Moving uses up charge, at a rate that is proportional to the distance moved, but depends also on speed and whether a baby is being carried or not. Holding an ogre uses up charge fairly rapidly.
If totally run down a nanny needs to be moved by two other nannies to a re-charging point. Recharging takes a certain time but can be interrupted if the nanny is needed for a task and has enough charge.
The nursemaid can always ask a nanny how much charge is left. When a nanny's charge gets below a certain level it spontaneously starts sending signals to the nursemaid. The intensity and frequency of the signals increases as the charge gets lower. Whether a signal gets through depends on its intensity and what the nursemaid is doing.
After babies have been in the area for a certain length of time they are taken away. New ones arrive from time to time.
The babies, nannies, snatchers and ogres have slight individual variations. Not all the entities in a given class move at exactly the same speed - some have faster top speeds than average some slower, though they don't always move at their own top speed. Some babies are more likely than others to wander towards ditches. As babies grow older their top speed increases, but some become more likely to go towards ditches others less likely.
Babies can't all be herded into the same place for safe-keeping. Any baby that is in the same rectangle as more than three others for longer than a certain time develops a thug-like tendency to damage other babies. Once this has developed the only cure is being left alone in a rectangle for a certain time. Occasionally the arrival of new babies can increase the population enough to make it impossible to ensure that there are no more than four per rectangle.
Protection of as many babies as possible until they are old enough to leave is the main goal of the nursemaid.
The nannies have no value in themselves, but are required for protection of the babies.
Faster babies are more valuable than slower ones, and older babies more than younger ones. Ex-thug babies are a less valuable export than wholly innocent ones.
This scenario can be indefinitely complicated, if necessary.
A possible intermediate deliverable might be a computer game representing this scenario with an interface for a human being to play the role of the nursemaid, receiving signals, formulating plans, and sending instructions. If so, players who become skilled at the game could provide "informants" for a knowledge engineering exercise.
Last Updated: 8 May 2004; 31 Aug 2014 (re-formatted.)
Maintained by Aaron Sloman
HTML version derived from plain text version 8 May 2004