Mapping Into Abstraction Such a Territory

The other week, S was reading one of the Stuart MacBride Oldcastle novels. Oldcastle is a fictional city in Scotland. The book she was reading had an OS-type street map of Oldcastle. So where the heck did that come from? This isn't just a map on a napkin or even the kind of thing you might find in an RPG supplement. (There are many wonderful things about White Dwarf 42 and one of them is that inckudes the first part of Irillian.) Someone has to put quite a lot of effort in with the mapping software. Which would be easier if tey had taken an existing city map and just flipped it somehow and then renamed the streets, buildings and districts. This is what I suspect was done as it would be the easiest thing to do, but I don't which city map they have used as the basis. If Oldcastle is "supposed" to be Dundee, you can't just flip the Dundee map as Oldcastle, like Dundee, is on the north of the estuary. Newcastle might be a possibility, of course.

The Old Devils is a very topographical novel. But is it just the topography of Swansea and the Gower with which Amis was very familiar or did he make it up?

When I was a child, one of my concerns with both plans of schools and maps of towns was missing something out. This couldn't possibly be a real school or a real town because a real school or a real school of its type would have these certain features that I have neglected. Getting hold of plans of schools in the early 1980s was not easy, but maps of towns were a bit easier. One could study enough maps of towns and produce a plausible enough looking map of fictional town I should think and perhaps that is what MacBride or someone in the design department at his publisher did.

But we know have generative adversarial networks. And, yes, at least one group has used GANs to produce images of imaginary cities. What is happening here is generating satellite images of Venice in the style of Los Angeles and vice versa and then using the filters to generate the imagesd based on handdrawn maps. You could you could then produce a map based on the satellite image. This doesn't help with missing out the police station.

There are hundreds of towns and cities in the UK. One could imagine using a GAN to generate imaginary maps that are difficuklt to distinguiish from thre real thing. The simplest system would just look at the road layout (plus indeed, infrastructure like railways, canals and rivers), but more sophiticated ones could look at the buildings and even the names of the streets. Yes, you might want some hand annotation, but I suspect you could generate reasonably convincing maps.

Of course, getting access to the maps might be no easy thing. Yes, you could have a go with Google Maps, but really you would want pretty full GIS data, so you are not using an image of the town. And some of this might require something closer to the solution of the general AI problem. How does the system know how many primary and secondary schools there should be or where the teritary college should be or the hospitals. How much knowledge of the world does it need and how much can it do just on the basis of features extracted from maps? Does it help know that a town the size of Preston will have a large hospital or can you just learn that with knowing anything about hospitals.

It's like the old Indonesia problem. Indonesia is a country in South East Asia. Malaysia is also a country in South East Asia. Singapore is a country in South East Asia, but it also a city. So Indonesia is like Malaysia is like Singapore. But Singapore is like Jakarta is like Kuala Lumpar. As countries, Indonesia and Malaysia are a lot more like one another than they are like a city state such as Singapore. But how do I learn this? As a person I have a concept of South East Asia, countries, types of countries, cities, capital cities, modern megacities in Rising Asia. But where do I get these categories from? How do I build a system that can learn these categories without having to be told them explicitly. There are knowledge representation systems such as a Cyc or Wikipedia, but what's the the state of the art in knowledge engineering. I live in hope that we will see a Coursera course on this one day.

It's not just fictional city maps that GANs could help with. There are also fictional languages. I will talk more about imaginary languages later, but GANs might be a great way to generate the kind of plausible Indo-European language I dreamt of as a child.

Comments

[John B]

I rather like this city generator, as they donn't skimp on the fortifications

https://watabou.itch.io/medieval-fantasy-city-generator

[Paul M. Cray]

That looks pretty good. Wonder if their maps would pass the GAN test. I am also interested in software that generates interesting maps of planets. That must be around somewhere.