Rob, I am trying to make improvements on the analysis of logical statements. My idea is that a program can analyze Satisfiabilty statements, (logical statements where the states of some logical variables are not assigned values), by using a system of selective compression. The problem of finding logical values (true or false) that can satisfy a statement can become very unwieldy in some cases. It is easy to find solutions to a logical statement like, A V B ^C. For instance A=T, B=F, and C=T will satisfy the statement. But with more logical variables (or literals), finding solutions to some logical satisfiability statements can become very complicated and time consuming. If an analytical program could operate on the compressed data (derived from a satisfiability statement) it might be able to find solutions to those statements faster than can be done at this time.
Logic can be used to express many kinds of significant relations in computer programs, so advances in finding solutions to these problems could yield great increases in speeds and memory efficiency (because logical satisfiability statements act just like compressions). Since I am talking about selective compression then if I could make any kind of advance it would mean that I could make some general advances in the field by showing that you can use selective strategies in compression (by using auto-generation of ad hoc sub programs which could operate with other individuated ad hoc sub programs). I was not really talking about lossy compression, but I did touch on that subject. An advance in Satisfiability could be leveraged in lossy compression. And as I was writing in this thread I started thinking about the somewhat mysterious discrepancy between human recognition and recall. I did not want to call it lossy compression exactly but if an agi program was able to create components of recognition that relied on ranges of values then by checking some input (the data from an ongoing event) against a system of these components to see if there was a fit that ability might explain the discrepancies between recognition and recall. The components of the system would not necessarily be called lossy compression because it might not make sense to try to decompress them. (So they would not serve as methods to recall some event.) You would only use them to see if an input could be fitted into them. (The system would also need some discrimination methods as well.) The system would need some way to check all possible combinations that might satisfy a given input so such a system would have some of the earmarks of a satisfiability problem. That is why I thought about this particular issue while writing this thread. One other thing about this. This has been heavily studied using fuzzy logic, neural networks and graph theory using weighted values (like probability networks). However, most of those efforts have not taken discrete concepts into account and they have not resolved the problem of using different kinds of weighted values (what I consider to be pre-Cauchy thinking based on an analogy from the contemporary miasma of using weighted reasoning today to the unexplained failures in using calculus before Cauchy.) Discrete concepts are important in recognition problems (using lossy methods if you want to use that term) because they are needed to find other related knowledge about the event that is occurring. However, this then becomes a satisfiability problem. So even I were to be a modern day Cauchy (I don't have the time to even pursue this fantasy) who could solve contemporary dilemmas in weighted reasoning, it would not be useful without some advances in satisfiabiilty problems. (I am planning on using Cauchy-appreciation in my logical satisfiabiilty analysis.) I hope you can get past all my off-track tangential thinking and get what I am saying as it relates to what you were saying. Jim Bromer On Wed, Jun 7, 2017 at 12:30 AM, Rob Freeman <[email protected]> wrote: > > Jim, > > It is hard to understand what you are suggesting. But I get a hunch you may be saying something like the idea I have been promoting for a long time. This is that for cognition all or most compressions will at best be partial (lossy?) > > I suppose you could keep a list of outlier data for each partial compression, and adjust processing based on it. > > Is that what you are suggesting? > > My guess is that we won't gain that much by compressing at the end of the day anyway. Processing distributed data in parallel will probably be as efficient as processing compressed data in serial. > > So my guess is it will be better to just work with uncompressed data, in parallel. > > But I like the insight that all compressions will be partial. If that is what you are suggesting. If that is what you are suggesting I agree with you. > > In short, I think the best way to address the partial/subjective compression problem I believe you to be addressing, will be ad-hoc compressions relevant to specific decisions. > > -Rob > > On Wed, Jun 7, 2017 at 3:16 AM, Jim Bromer <[email protected]> wrote: >> >> A program could create new rules at a higher abstraction level that could operate on the rules that exist at a lower level. (This characterization is relative but I think it is useful.) This might allow for greater compressive individuation. >> >> So creating individuated sub programs based on the data that was being compressed (or transformed from one compressed form to another) might be useful in devising ways to effectively work with the data without fully decompressing it. >> >> Logic is (typically) a compressed representation and you can often work with in compressed form. But, there are many common situations where that data has to be excessively decompressed in order to work with it. My conjectured-goal is for a program to be able to create new rules during an analytical run so that it could individuate the data (to compress it) and still be able to operate on the different individuations without fully decompressing them.Even if I can’t get it to work, I should be able to show that it can work on some special cases. >> >> >> Jim Bromer >> >> On Tue, Jun 6, 2017 at 10:02 AM, Jim Bromer <[email protected]> wrote: >>> >>> I meant: >>> This would allow for greater individuation which is necessary in order to work with the data in compressed form because the data would be too extensive for the program to handle it in decompressed form. >>> >>> Jim Bromer >>> >>> On Tue, Jun 6, 2017 at 10:00 AM, Jim Bromer <[email protected]> wrote: >>>> >>>> Most programs operate on a higher level of rules which operate on the data that is encountered. This higher level program is a kind of general abstraction. A program can create new rules at higher abstraction that could operate on the rules at a lower level. (This characterization is relative but I think it is useful. Programming languages do this but the essential part of the program produced is created by a person) This would allow for greater individuation which is necessary in order to work with the data in decompressed form because the data would be too extensive for the program to handle it in decompressed form. This individuation would mean that the data could not be decompressed without the higher level abstractions (or sub programs) that each individual run would produce. These higher level sub-programs would be part of the data the program would create but I am trying to emphasize the idea that the program would create sub-programs based on the data that it was encountering and compressing. (My idea of working with SAT is that it would transform statements in traditional Boolean Logical form into another form which would also be compressed.) >>>> So creating individuated sub programs based on the data that was being compressed (or transformed from one compressed form to another) might be useful in devising ways to compress the data. >>>> This kind of program would use generalizations as components both to directly represent the data in compressed form and to create sub-programs that could operate on the compressed data. >>>> >>>> Jim Bromer ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
