Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!gatech!bloom-beacon.mit.edu!panix!zip.eecs.umich.edu!umn.edu!news
From: "James Albert Larson" <larso171@maroon.tc.umn.edu>
Subject: Re: Q: Microsoft Excel 5 spreadsheet on inter
To: edguzis@aol.com
Message-ID: <38417.larso171@maroon.tc.umn.edu>
X-Minuet-Version: Minuet1.0_Beta_16
Sender: news@news.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: dialup-4-15.gw.umn.edu
X-Popmail-Charset: English
Organization: University of Minnesota, Twin Cities
Date: Mon, 29 May 1995 12:37:54 GMT
Lines: 80

On Sun, 28 May 1995 13:51:26 -0400, 
edguzis@aol.com  <edguzis@aol.com> wrote:
>Jim,  I would be interested in looking at your Excel 5 application when
>you get it posted somewhere.  

Hi Ed, thanks for your interest.  I'm posting this letter to the newsgroup 
as well in case anyone else is interested.

The Excel spreadsheet models and "optimally" solves the electric utility 
generation scheduling problem, known as UNIT COMMITMENT (UC).  Given 
a planning horizon of 120 hours, and 25 generating units that can be 
scheduled, it find the minimum cost generation schedule.  A generation 
schedule is a 25 X 120 matrix of bits, where matrix[i,j] = 0 means that 
the unit i in hour j is off, and matrix[i,j] == 1 means that unit i in hour 
j is on ).

Throughout, "optimally" and "optimum" are in quotes, because there are 
2^(25 X 120) possible generating schedules, and finding the pure perfect 
optimum is impossible with today's and tommorow's computing technology.

Thee current near-ready version of this Excel 5 application has 
nothing to do with GA.  Not yet, and not for quite awhile.  It uses a 
traditional dynamic programming (DP) technique exclusively.  Also, the 
manual, although lengthy (70 pages), does not really explain the UC 
problem.  Its more of a user manual for people who already have some 
background in the UC problem.  Therefore, this version will be uploaded (if 
at all) to a UC researcher's directory.  Ed, I'll let you know where it 
ends up.

The value of the DP technique is that it finds the guaranteed perfectly 
optimal (optimal not in quotes here) solution to a *small* scale sized 
generation scheduling UC problem with a couple of constraints ignored.  
As such, it provides a research tool for studying the characteristics of 
perfectly optimal solutions to this problem (or more accurately to the 
modified and smaller - scaled version of the problem).  As for what small 
scale is, the current Excel spreadsheet handles 8 units without excessive 
disk thrashing. 

The problem of the dynamic programming approach is the curse of 
dimensionality.  The problem size doubles with each additional unit.

Currently, a lagrangian relaxation approach is gaining favor in the 
industry as the best approach to solving the UC problems of medium 
and large utilities.  But it doesn't provide the optimal solution, and its 
run time is not all that fast either.

The new kid on the block is a GA solution to it.  I've had some success 
with it, and have ideas to greatly improve on it.  I believe GA is the only 
approach that has potential to handle some real tough flavors of the 
UC problem, such as those involving purchases and sales to other 
utilities (very important but poorly handled by commercial UC programs), 
fuel availability constraints, pollution emissions, and on and on.  
Commercial UC programs handle these heuristically and poorly (or not at 
all), forcing the dispatchers to "fix" the solution manually.

Since the current Excel spreadsheet only uses dynamic programming and no 
GA, I wouldn't be uploading it to a GA repository, or pestering the GA 
community with it.  But when I get some kind of GA version put together, 
then I would probably upload both versions.  The DP version, by providing a 
purely optimal solution to a small-scale modified version of the problem, 
would be a training tool in showing what guaranteed optimal solutions look 
like, and as a benchmark to the GA version.  And of course, the manual 
would cover the UC problem from square one.

An uploadable and documented GA version is at least 2 months off.  It will 
not be particularly brilliant.  It will rely heavily on heuristic genetic 
operators, and probably will be considered a hybrid scheme.  My main 
problem with combining heuristics and GA is that there is a big danger that 
my heuristics will drive the solution to what *I* think the optimal 
solution is, and not allow the GA full freedom to find its own 
potentially better optimal solution.

My real goal is to produce a version that combines the best of DP, 
Lagrangian Relaxation and GA.  I only have some faint conceptions of how 
that might be done.

Sorry, I only meant this to be a couple paragraphs.

Jim Larson

