Applying AI to classify the ships of Oolite

[Makandal]

Some might remember this thread we had about one week ago. I suggested to use AI algorithms to get a classfication of the ships of Oolite. However, it necessitated to enter ships selected characteristics in an excel spreadsheet.
I am stuck in an hotel in Accra, I cannot decently go to the beach before the closing of work and I promised as well to be faithful to my very beautiful fiancee. In summary, I am bored.
So I started this tedious job.
And I build a decent workflow to analyze the ships.
I just want to tell you that the results are surprizing and after reflexion, very logic. I still have to enter the oxps ships from G to Z.
Presently, my workflow found various ways of classifying the ships. One is interesting, the one of an algorithms called hierarchical cluster analysis (HCA).
It found 6 classes:
- The very light vessels: Moray Medical, Adder, Cobra Mk1
- The fast, light armed vessels: aphid, dragster, firewasp....
- The multipurpose medium vessels: boa, python, cobra, fer de lance, drake etc...
- The multipurpose light vessels: asp, firefly, excalibur, some eagle
- The immortals: assassin, caduceus (high recharge rate)
- The Anaconda: alone, in its glorious majesty.

Some classifiers (k-N clustering for example) found other classes almost the same as well.
I decided to use also a C4.5 classification tree because it is usually a very powerful way to enhance classes. It creates 5 classes using only the speed (>0.312>), the size and the possibility to add a military shield.
.
I will keep everyone posted and send you the results when I finish.

[Makandal]

This is the very first result with a complete family tree of the ships of Oolite.


http://www.flickr.com/photos/59450711@N08/5516913815/

[Malcul]

Sorry Makandal, I've looked at your resulting sheet, and can make nothing of it.

You have seven classes on display with no labels. There are what appear to be elongated boxes alongside the various ships, do they mean anything?
I love my own craft, the falcon, and I cannot see any way of comparison with anything else, except a colour or clasification.

I don't do maths or statistics, but I thought I was only ordinary stupid, not really, really stupid.

Please explain your work, I am interested. I love graphic displays of information, easier to remember!

(If you are really bored, you could try a falcon and tell me why I'm missing something!)

[Mauiby de Fug]

Cool! A little hard to see, and a bunch of ships I don't know the stats of particularly well, but it seems to have grouped them well. A lot of ships I think of as similar are "related", so it seems to be accurate enough to me! As I thought it would be: interesting!

Malcul - I'm working on the assumption that the different colours correspond roughly to the classes mentioned in Makandal's first post about he preliminary results. That said, I agree that some labels would definately be helpful!

[Smivs]

I have to admit I'm baffled by this as well.
Some sort of stated classification would have been nice, as I see ships together (the colour groups) which I don't see as being remotely similar, and lots of boxes of various length and width with no way of interpreting what they mean. Is long good, or is wide good or what?
Also, I don't see the 'benchmark' core uber-ship the Constrictor (or it's derivatives) anywhere, so a major reference point is absent (unless I've gone blind, which is possible ).

[Gimi]

I think this is more of a technology demonstration. It did give me some information when I compared the different ships that are grouped together. However, for this to be useful, the actual chart needs to have a lot more information, including an explanation of colours, lines and what parameters are used to do the classification and what weighting they have been given if any..

[Malcul]

Mauiby de Fug, interesting it certainly is, and if I do like you and take the classifications as listed in Mak's first post, then my falcon is an immortal!?! It is good, ... but the pilot lets it down!

I am looking forward to some conclusions from Mak's work, I would like to see some suggestions for some under-rated craft in various categories with some reasons. Hell's teeth, that sounds like an exam question!

If you're slow, you're dead.
If you have no carrying capacity, you've no credits.
If you are big, you are an easy target.
If you are painted bright yellow, they can spot you a mile/kilometer/Ly off.

... Taxi!!!

[Zireael]

The lines grow up from the ships (rotate the image 90 degrees right and you'll see. The closer the ships are related, the closer the lines are. And the colours show where one wide class is. Similar ships are together (two Vampires, two Imperials etc) and the farther the ships in one class are, the less similar they are.

I'd classify the yellow class as 'immortals', class light blue as 'ultimate combat machine' and class orange as 'all-rounders'.

[lohwengk]

Looks interesting, although I'm not sure how to make use of the information.

Anyway, will the ship price be included in further analysis?

How about factors affecting combat performance, e.g. roll, pitch, yaw, ship energy, energy recharge rate, thrust, ability to mount witchdrive injectors, etc.?

[Makandal]

Sorry, I dropped the file in a hurry, I was leaving the hotel in Accra to fly back to Lagos.
Ok, let me give more details. This is not a 'classical stats' classification. This is multiparameter classification and it gives a kind of family tree of the ships by their similarities.
The parameters are the following: size (lxLxh), cargo capacity, extra cargo bay capacity, max speed, maneuvrability (roll x pitch), nb of guns, nb of missile slots, shield booster available, military shield available, price, vanilla Oolite ship or not. About 3% of the information where missing, so I have to infere it.
This classification does not put a priority on any of the parameters. I will join the same analysis on the parameters to show you how they are related. This is also very interesting, let me analyze it:
- Cargo capacity and size of the ship are correlated (positive correlation in that case): logic
- Speed, energy recharge rate and price are also strongly correlated.
- The maneuvrability and the nb of missile slots are also correlated.
- Finally, gun mounts and shields are closely correlated as well.
- On more lose basis, we see that the first 3 set of items are related while the other one (Guns & shields) is independant.

Important: correlation is different from causation.

You can look at the same for the other tree of ships which is much more complex
The colours just correspond to some cut off I decided to set.

I want to state that this is one the results. I have other plot which is multidimensional scaling (MDS). Here you can see another way of clustering the ships together. It's not easy to read the name. For those interested, I can zoom on specific areas.

http://www.flickr.com/photos/59450711@N08/5519691612/

http://www.flickr.com/photos/59450711@N08/5519691380/


EDIT: Finally, I have a last analysis which can be very useful, this is called Naive Bayes Analysis. I will describe it later, need to finish to cook the Spaghetti Carbonara for my daughter and her friend.

[CheeseRedux]

price, vanilla Oolite ship or not.

The way I see it, these two are not particularly useful, especially when

This classification does not put a priority on any of the parameters.

Theoretically, two identical ships with a very different price tag could end up being classified as being further apart than two very different ships with identical price tags.
Price has a very limited function; it affects whether you can afford the ship or not.
One potential use for your classification is for people to look for similar ships in order to get a poor-man's-version of their dream ship if they cannot afford the real thing. Having ships classified as more or less similar based on price hinders that.

And OXP/Vanilla? That one is quite silly.

[Gimi]

Actually what I might want to use this for is selecting ship OXP's to install. If we use the group option, and then use parameters that group ships that are weighted in performance together I can see what I need to be careful about installing in order to keep the game balance.

E.g. High speed, good manoeuvrability, huge size and big cargo hold would fall outside the average ship group and then be identifiable.

[Zireael]

I think the price is not really relevant, especially in OXP ships. Some are grossly underpriced... or overpriced.

[Commander McLane]

Actually what I might want to use this for is selecting ship OXP's to install.

Many people would select their ship OXPs much more by the graphics, which don't play any role at all in this grouping.

[Gimi]

Actually what I might want to use this for is selecting ship OXP's to install.

Many people would select their ship OXPs much more by the graphics, which don't play any role at all in this grouping.

I agree, but there are some rally pretty "über" ships out there and, while graphics are important, playability and consistency means more to me. Hence, the Super Cobra is not in my Oolite (There was only one and it's Über), the Cobra Currier is but I'm waiting for the Griff_Player_Multi_decal_scuffed_cobra_courier.oxp with anticipation.