Is optimisation so bad?

27 07 2009

Tools are not bad, the way we use them may be. Let’s first take a look at the question: What is optimisation?

If we seek for a general definition of optimisation, this could be “to provide the best possible answer to a question according to a criterion.” What are the ingredients of this definition?

  1. Question
  2. Answer
  3. Criterion
  4. Optimisation method

The answer (2) depends on (1), (3) and (4) above. It is possible for (2) to be different (but not imperative) if you change (1), (3) and (4). Therefore optimisation per se is not a good or bad practice but merely a way of answering questions. For it to provide a reasonable answer, you need to use an appropriate criterion and method to a “well posed” question. “Well posed” is rather a motivation to make sure the question you ask is the one you really care about, i.e. it is sufficient to describe the problem you have in hand.

In a more mathematical manner, we could describe optimisation as follows:

“within a given range, find the extremum of a function, where the extremum satisfies some requirements”.  (a)

This implies that

  1. There is a function that describes our problem
  2. There is an extremum within the given range

For an analytical function, the extremum can be accurately calculated analytically.

For a numerical function an approximation can be given using numerical methods.

Such a numerical optimisation needs additional input, aka the accuracy to which the extremum is to be sought. Depending on the method used, the accuracy can be treated in a different manner.

Now, most problems can be translated to the statement (a). However, it is important to know what kind of problem one needs to solve and not confuse the definitions. Let’s take an example where equally interesting questions can yield different results that under some circumstances may be confused with each other.

Let’s suppose you’re standing on the edge of a cliff over a gorge and you want to get to the bottom of that gorge. This is a rather vague need and we might stipulate the following problems:

  1. Find the deepest spot in the gorge
  2. Find a sufficiently deep spot in the gorge
  3. Find the fastest/slowest/safest way to the deepest spot
  4. Find the fastest/slowest/safest way to a sufficiently deep spot
  5. If the spot you want to visit is known (you have seen it in a book) and maybe you even know where it is, find the fastest/slowest/safest way to that spot.
  6. Find a spot which is deep enough, but not that deep so that you can climb back up.
  7. Find a spot which is deep enough but on a bump where you can stand and take pictures, which bump is not that steep so you totter.
  8. If a friend is where you want to be, and he can see you with his binoculars but you cannot see him, you might want him to come and get you rather than you try to find him.

In some situations above a “sufficiently deep spot” is sought. A sufficiently good solution is sometimes the best solution to a problem, as it provides a good balance between the result and the effort to reach that result. For example let’s say you’re using the Navier-Stokes equations to choose which shape has minimum drag and maximum lift between a circle, a square, a line and an airfoil and perturbations of these. Any airfoil is a sufficiently good solution.

The CAE case

Computer Aided Engineering means using a numerical system to simulate a real system and take decisions based on the outcome of tests performed on the numerical system. One can only hope the numerical system simulates the actual system sufficiently well.

Assuming the numerical system can be a good enough approximation, one has to compose the right system for his problem. Considering the case of the airfoil, a designer could stop when he finds that among the attainable shapes, the airfoil is the one that provides more lift for the least drag. But he could go further and optimise the shape of the airfoil for a particular speed range and angle (or other service requirements). The result might be a quite peculiar cross-section. Let’s say that we construct one such cross-section and test it, and let’s assume that the results of the tests are identical to the simulation results. In this ideal case, the numerical or virtual system acts like a gauge to the real or physical system. But in order to achieve this result, a high precision manufacturing process was needed, for which the time length and cost of it might not be possible to justify.

The designer, answered the question

“Which is the airfoil that performs best under the given service conditions”?

The assumption here is that if the airfoil must be constructed, there is potentially infinite time and funding for trial and error until the result conforms with the simulation, given that the numerical system accurately simulates reality.

If a series of airfoils are to be manufactured though, a less time consuming and costly method is required. This additional restriction results in departure from the ideal case and landing into the reality of

System + measurement = variation

In this situation the “optimum” might change, which is normal, because manufacturing is a different problem to designing.

Now the question is

“Which is the airfoil that performs best under the given service conditions and can be manufactured with the particular process”?

Obviously this is a different problem. The system is no more the airfoil alone but the manufacturing is introduced as well. But this is still an optimisation problem.

So identifying the problem you actually want to solve is a first step to getting it solved in a satisfactory manner. Obvious statement but often ignored. Take another example: consider the case of a metal forming process, i.e. forming a part from a coil, and try to look for the optimal thickness of the coil so as to minimise weight. The possible coil thicknesses are 0.5, 0.6 and 0.7 mm. Let’s say a designer performs a simulation for each case and no cracks are identified for any of these. Let’s assume that an optimisation algorithm is also used which identifies 0.500 as the optimal thickness, with 0.499 giving cracks.  The designer would pick 0.5mm as the optimum thickness. So far so good, for what we asked we got an answer, there’s no statement here that we actually want to produce this part. Now, if we go to production, we might find out that the manufacturer gives us a variation of 0.01 for the coil thickness, which will result to scrap. Assuming that the manufacturer gives the same variation for all coils, the production engineer would pick 0.6 as the optimum thickness that eliminates cracks and minimises weight. Is this the best solution? It is good enough for the question asked. If the coil supplier is asked to provide statistical data in terms of a median and standard deviation for his coil we might calculate a scrap rate of a% for the thickness of 0.5. Then we should see if this scrap rate is acceptable and try to optimise the cost of production if we want to reduce it. That’s yet a different problem.

So, all in all, optimisation is a method for seeking the best solution for a given problem. As with all tools, one has to know how to use it. It is not bad or good. However, sometimes it is misused.

Advertisements

Actions

Information

3 responses

27 07 2009
erickam

as you say “optimisation is a method for seeking the best solution for a given problem”. The given problem is where many CAE issues go astray.

when the problem is placed before CAE engineers to “find a solution” they will invariably find one.

When we ask “is the solution feasible?” the answer is “YES” but the fine print we often miss is under what conditions this “feasible” solution exists.

27 07 2009
dpservis

Eric

you’re right, but you might as well consider that we are asking the wrong question. What I try to show here, is that any question can fit this general definition. You just have to be very careful when defining the question. For example, if you need to manufacture a special component for the international space station, you might as well choose a technically optimal design and manufacture it at any cost. This is of course not the case in mass production. There, manufacturability requirements must be introduced. But still this is an optimisation problem, only for a different system. In our practice, selecting the most robust solution is still an optimisation, even if heuristic criteria are used.

Optimisation is a way of solving problems and one way of optimisation (actually the simplest) is to minimize or maximize an objective function. I think your argumentation is correct in that people confuse the latter with the former and in this respect we should not use the term “optimisation” to imply such a general context if people cannot understand it.

28 07 2009
erickam

D,

“you might as well choose a technically optimal design and manufacture it at any cost” …. there you’ve hit my biggest pet problem on the head!

How do we get the “solver snobs”, “Mesh Monkeys”, “CAE gurus”, and “PhD-Pile it Higher and Deepers” of the world on board for that one additional metric for the objective function. We need to be able to afford the “it”. Whether we make one or One Million.

For most CAE problems we can define a limit (failure limit, fatigue limit, intrusion envelope, entropy state) but how do we do so for things like Cost, or Time-to-Delivery, Surface Quality? And once we do so “how do we compare the opportunity cost for pursuing one over the other. How much of a loss in one objective function is worth gaining a little more in another?

(D, this is a little tiring typing I’ll be in Zurich in a few weeks we can talk more over beer)

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s




%d bloggers like this: