Re: on optimization, again

At 08:35 AM 18/10/2002, Pedro wrote:
>I would also like to relate these adaptive games to the processing and
>self-constructing limitations of the informational agents (so the need of
>specialization and the further problems of integration found in all
>realms: molecular, neuronal, social, interdisciplinary). And also to the
>considerations that John H made weeks ago in his comments to Floridi: 'Is
>the flow of information quantitatively directional?' I believe it is, and
>that the above processing march by the info agents under the guidance of
>the invisible hand towards their natural places (necessarily including
>'extinction, degradation, etc.) impinges into our own conversion
>data/meanings too. Meaning would be best understood as a byproduct of
>ontogenetic adaptation...

Some caveats: 1) As Elliott Sober notes in Reconstructing the Past, if
adaptive processes alone determine final state, then information about the
past is lost. In statistical mechanics, equilibrium means that information
about the microstate of the system is minimized. Although the Shannon
information of message is maximal when the messages equiprobable, knowledge
only that the state of the system is such that the messages are
equiprobable minimizes the amount of information about messages relative to
that given by other distributions. The general rule is that movement to
steady state (including equilibrium) means a loss of information, both
about state structure and past dynamics. From this I conclude that we also
need to look at information generation and preservation rules. 2) although
systems mover towards attractors, I think that it is rare in nature to find
them in attractors, so the information in states that are non-optimal may
be of more practical significance than the information determined by
optimality conditions. 3) even if optimal states are obtained, the
information about the route to optimality is not retained (see point 1). It
may well be that the optimality is of less significance than how it is
attained. An example comes from the Shannon message example in (1): the
optimality of the message gives minimal information about the specific
message; this is given by the way the message is formed, which optimality
does not address.

These three issues underlie my recent scepticism about optimality methods
and their value to information theory. On the other hand, I see the local
approach towards optimality (with possible changing endpoint due to
non-linear interactions) as a major part of the dynamic of information. I
see this as a generalization of the Principle of Least Action, which I
mentioned in a previous post.

John
Received on Sat Oct 19 01:48:06 2002