Features of the UCI chess data sets

Only to be found in Figure 2 of CONSCIOUSNESS AS AN ENGINEERING ISSUE, PART 2 by Donald Michie (University of Edinburgh).

1	bkblk	the BK is not in the way
2	bknwy	the BK is not in the BR's way
3	bkon8	the BK is on rank 8 in a position to aid the BR
4	bkona	the BK is on file A in a position to aid the BR
5	bkspr	the BK can support the BR
6	bkxbq	the BK is not attacked in some way by the pro- moted WP
7	bkxcr	the BK can attack the critical square (b7)
8	bkxwp	the BK can attack the WP
9	blxwp	B attacks the WP (BR in direction x = -1 only)
10	bxqsq	one or more Black pieces control the queening square
11	cntxt	the WK is on an edge and not on a8
12	dsopp	the kings are in normal opposition
13	dwipd	the WK distance to intersect point is too great
14	hdchk	there is a good delay because there is a hidden check
15	katri	the BK controls the intersect point
16	mulch	B can renew the check to good advantage
17	qxmsq	the mating square is attacked in some way by the promoted WP
18	r2ar8	the BR does not have safe access to file A or rank 8
19	reskd	the WK can be reskewered via a delayed skewer
20	reskr	the BR alone can renew the skewer threat
21	rimmx	the BR can be captured safely
22	rkxwp	the BR bears on the WP (direction x = -1 only)
23	rxmsq	the BR attacks a mating square safely
24	simpl	a very simple pattern applies
25	skach	the WK can be skewered after one or more checks
26	skewr	there is a potential skewer as opposed to fork
27	skrxp	the BR can achieve a skewer or the BK attacks the WP
28	spcop	there is a special opposition pattern present
29	stlmt	the WK is in stalemate
30	thrsk	there is a skewer threat lurking
31	wkcti	the WK cannot control the intersect point
32	wkna8	the WK is on square a8
33	wknck	the WK is in check
34	wkovl	the WK is overloaded
35	wkpos	the WK is in a potential skewer position
36	wtoeg	the WK is one away from the relevant edge

References:

1. UCI archive


2. PDF of the Michie article, found on Google

Visualizing clusters of high-dimensional data

I wanted to demonstrate that distance-based methods should have a hard time distinguishing some classes (say the libras UCI data set) and not others (say the iris UCI data set) with a visualization. S.R. suggested:

1. Use multidimensional scaling (MDS@wikipedia) to plot points in 2 dimensions.
   
2. draw lines between points and their nearest neighbors in the original space

If the lines are an overlapping mess, the points don't cluster well.

Steps to do this in MATLAB:

1. M=dlmread('my.data', '\t') % read tab-delimited data into memory


2. M=transpose(M) % I have to transpose my data to make columns examples, rows are variables (features)


3. D=dist(M) % compute distances into a square matrix


4. G=mdscale(D, 2) % find points in 2 dimensions

G is now a Nx2 matrix (N=number of samples). The rest of the process involves plotting the lines to connect neighbors. You can do that in MATLAB, but I don't know a particularly fast way.