Hi Arnoud, Two points:
1. Most real environments are open, non-linear systems, which means they generally have some characteristic time horizon beyond which prediction is no better than a guess based on historical statistics, no matter how much computation is used. For the weather, for example, predictions beyond two or three weeks will never be much better than just using climatological averages. 2. How well an agent has to do is often relative to other competing agents, or when there is no competition all an agent can expect is "as good as possible". For example, in predicting the stock market an agent only has to be better than competing agents to make money. And in predicting the weather, there is a real limit on how well an agent can do. Cheers, Bill ---------------------------------------------------------- Bill Hibbard, SSEC, 1225 W. Dayton St., Madison, WI 53706 [EMAIL PROTECTED] 608-263-4427 fax: 608-263-6738 http://www.ssec.wisc.edu/~billh/vis.html On Thu, 18 Sep 2003, arnoud wrote: > Dear AGI, > > What is the maximal complexity of an environment in which prediction of future > events in the environment is still computationally feasible? > > In most 'realistic' environments prediction of the near future needs to be > very precise and prediction of the far future can be vaguer, i.e. a large > class of event types satisfies the prediction, in order for an agi agent to > achieve its goals. (If prediction on the very long term also needs to be > detailed and precise the environment is impossible for any agent, not?). > > In a formula, : > (maximal vagueness of prediction that is allowed by agi agent, in order to > plan and act successful to achieve goals) / (time scale of horizon of > prediction) = some constant. > > Vagueness of prediction is the number of perception event types that satify > the prediction. (count the perception event types at bit level). > > somewhat equivalent formula: > (number of patterns of about time length l that occur in the environment) / > (about l) = some constant > > A pattern here is a class of noisy variations on a pattern (10% noise, 20% > noise? The smaller the constant gets the higher the noise ratio can be.). > > How large can those constants be? How complex may the environment be maximally > for an ideal, but still realistic, agi agent (thus not a solomonof or AIXI > agent) to be still succesful? Does somebody know how to calculate (and > formalise) this? > > Bye, > Arnoud > > > ------- > To unsubscribe, change your address, or temporarily deactivate your subscription, > please go to http://v2.listbox.com/member/[EMAIL PROTECTED] > > ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
