http://wiki.lesswrong.com/wiki/AIXI
So you could use the time loop to run through this algorithm.
But what's this about predictions? Why predict when you have a time machine? All you really need is a reward function and this algorithm:
- Get a message from the future with its best-ever reward and how it achieved it.
- Try something else
- When you get to the future, see whether you've done better than the best ever. If so, replace the best ever.
- With each iteration keep a count. If you've reached some maximum number of tries without improving the best score ever, do the best strategy one last time and quit.
It might be tricky figuring out the best span of time to use, and whether you can do mini-loops within bigger loops somehow.
Or try simulating a human brain.
As well as performing arbitrary computations in one tick, we could also keep a CTC going as a "reset button": whenever we realise that an incorrect/sub-optimal decision has been made, we can go back to when we started.
This lets us 'brute force' our real-world interactions (ie. our IO), as well as "merely" gaining an infinite computing speed-up.
This could be exploited by something similar to the "Success Story Algorithm" http://www.idsia.ch/~juergen/directsearch/node13.html . The CTC-AI would some way of "scoring" the world (ie. a fitness function), some probability of modifying its decision-making policy, and some non-zero probability of resetting the world. If the reset probability is anti-correlated to the world's score, worlds which score higher over time are more likely, hence the machine's decisions are more likely to increase the score than decrease it, hence the policy is more likely to improve over time.
But seriously, the problem with this is that it assumes that the only resolution to the CTC is that the AI generator tries another program.