Home Reinforcement Learning and Artificial Intelligence (RLAI)

The empirical knowledge hypothesis

--Rich Sutton, Sep 10 2004

The ambition of this web page is to state, refine, clarify and, most of all, promote discussion of, the following scientific hypothesis:

(All) world knowledge is translatable, without loss of meaning, into statements about, and comparable with, future lowest-level sensations and actions

and the slogan

Knowledge is prediction
return to hypotheses


World knowledge
is knowledge about the world, that is, particular to the world one finds oneself in.  Knowledge that would be true in any world, such as mathematical knowledge, is specifically excluded.

Sensations and actions are the lowest-level input and output signals available to the agent possessing the knowledge.  The agent is presumed to be situated in time, that is, with respect to a single time series of experience, of sensations and actions, the data of life, as discussed under the experiential AI hypothesis.

The parenthetical "(All)" is primarily a definitional issue.  It is meant to cover all commonsense world knowledge such as knowing the layout of your home, knowing your name, knowing the capital of Slovakia, knowing the shape of a wine glass, knowing that f=ma, knowing that the earth orbits the sun, etc.  As discussed above, mathematical knowledge is explicitly excluded.  A more subtle issue concerns knowledge that is specific to your world but not explicit, such as having a good set of features for your world.  One could consider this an exception to the hypothesis, or one could consider such generalization-assisting knowledge as also being a statement about future experience.  But this is a tricky fine point, and to focus on it would be to miss the main thrust of the hypothesis, which is about knowledge as we regularly think about it.

"translatable, without loss of meaning" means that the knowledge essentially is those statements about the future.  The knowledge is no more than the sum of those statements, and the sum of those statements is all there is to the knowledge.  The knowledge is not about a process that produces the data, except insofar as it is about the data itself.
--Rich Sutton, Sep 10 2004
The word "future" could be removed from the statement of the hypothesis, thus including memory of old sensations and actions as knowledge, while retaining most of the hypothesis' meaning and impact.  I personally have no problem with this, though I slightly prefer the given form because it is sharper.  But peace, this is a question of how we choose to use words, and either way is reasonable.
-RS --Rich Sutton, Sep 10 2004

Extend this Page   How to edit   Style   Subscribe   Notify   Suggest   Help   This open web page hosted at the University of Alberta.   Terms of use  6255/0