'die' is my own term, since it seemed to be the game term analogous to 'when an agent makes a move that renders it causally unconnected to all future rewards' (again, my own description).
The problem with including games in which one can 'die' is that they take much longer to learn. Suppose the agent the first time it plays the game happens to 'die', and now it only experiences a steady stream of 1,1,1,1,1... (low rewards). Nothing it does changes its future rewards, so exploration (trying different moves) is penalized. Dying on the first move might look like a good strategy!
Imagine if the rules looked like this: die~>1,1,1,1,...; not-die~>either -1 or +10. If the agent first tried out die, saw the +1 rewards, then the next game chose not-die and got -1, it may permanently start exploring down the die branch. An agent might eventually go back and try the not-die route and finally discover the +10, but this would take a while and is at odds with the idea of a reasonably quickly administered IQ test. Better to exclude such tests and switch to a more complex one.
Yes, now that I think about it, I guess their formalism tends towards incredibly low-signal environments where the actions are primarily simple "tokens" that can be named suggestively but aren't capable of actually revealing the data needed for the kind of sophistication I'm thinking of. That is, The environment is generally incapable of displaying an environmental tag that would suggest "novel action X (unlike novel actions Y or Z) could be dramatic and irreversible".
The only way to acquire such insight in a totally "from scratch...
"Measuring universal intelligence: Towards an anytime intelligence test"; abstract:
http://www.csse.monash.edu.au/~dld/Publications/HernandezOrallo+DoweArtificialIntelligenceJArticle.pdf
Example popular media coverage: http://www.sciencedaily.com/releases/2011/01/110127131122.htm
The group's homepage: http://users.dsic.upv.es/proy/anynt/
(There's an applet but it seems to be about constructing a simple agent and stepping through various environments, and no working IQ test.)
The basic idea, if you already know your AIXI*, is to start with simple programs** and then test the subject on increasingly harder ones. To save time, boring games such as random environments or one where the agent can 'die'*** are excluded and a few rules added to prevent gaming the test (by, say, deliberately failing on harder tests so as to be given only easy tests which one scores perfectly on) or take into account how slow or fast the subject makes predictions.
* apparently no good overviews of the whole topic AIXI but you could start at http://www.hutter1.net/ai/aixigentle.htm or http://www.hutter1.net/ai/uaibook.htm
** simple as defined by Kolmogorov complexity; since KC is uncomputable, one of the computable variants - which put bounds on resource usage - is used instead
*** make a mistake which turns any future rewards into fixed rewards with no connection to future actions