In the past we’ve discussed several sciences that rely on principles of intelligent design: among them, archaeology, forensics, cryptology, informatics, SETI and SEETI, and more. Now a paper describing a remarkable achievement offers the opportunity to explore another: optimization. The paper by University of Washington machine-learning researchers Abram Friesen and Pedro Domingos is well described in a UW news item. Jennifer Langston defines optimization and tells how useful the field is.
The key to solving many of the most important problems in business, science and technology lies in optimization — finding the values for variables that give you the highest benefit.
Whether those are which stocks to buy, which search results to return, what best predicts the outcome of the next presidential election or which amino acids to string together in a new drug to fight malaria or cancer, optimization is crucial to getting what we want. When a problem is simple, we can program a computer to solve it. When it’s too complex for that, optimization is how the computer finds the solution by itself. [Emphasis added.]
We all know a computer is incapable of finding a solution all by itself. An appropriate algorithm designed into it, however, can seemingly hit the jackpot automatically.
In essence, optimization is a search strategy. William Dembski deals extensively with search in his latest book, Being as Communion. Search is closely tied to the concept of information, which Dembski defines as a winnowing process that realizes possibilities by ruling out others. If the information desired is a solution with the highest benefit, a search must be conducted for that information. The science of optimization, therefore, is the study of algorithms or strategies that realizes the optimum solution by ruling out others. As such, it is an information-rich field.
Friesen and Domingos have scored an impressive achievement in this field. Their new optimization algorithm, named RDIS, performed an optimization task between 100,000 and 10 billion times more accurately, on average, than standard techniques. It was so good it won the top prize at the 24th International Joint Conference on Artificial Intelligence in July. This giant leap in optimization can help speed up finding solutions listed by Langston, and others like road recognition for self-driving cars and determining folds for custom proteins.
Domingos explains why optimization is more important than many people realize:
“In some ways optimization is the most important problem you’ve never heard of because it turns up in all areas of science, engineering and business. But a lot of optimization problems are extremely difficult to solve because they have a huge number of variables that interact in intricate ways,” said senior author Pedro Domingos, UW professor of computer science and engineering.
Figuring out protein folds, for instance, has been a long-standing problem. A protein begins with a chain of amino acids, but folds into a complex, three-dimensional shape. For all but small proteins, predicting the shape from the chain — given that the amino acids interact in complex ways — has defied many previous searches. RDIS may be able to speed the solution for synthetic drugs by optimizing the amino acid sequence for the desired shape. Another application is to determine a three-dimensional shape from a series of two-dimensional images.
Making RDIS work so well relied on several techniques. One was a divide-and-conquer strategy called decomposition. By breaking down the problem into smaller problems and treating them as independent problems (even though they are not), the designers converged on the optimal solution much faster. Another strategy is recursion. These and other technical details are described in the paper.
Optimization is not limited to AI or computers. Every logically thinking person faces optimization problems; e.g., “How shall I distribute my funds to optimize my return on investment?” “What players shall I call up for this football play?” “Shall I drop off the kids at school before going to the bank, or after?” “Where is traffic likely to be worse?” Optimization rapidly becomes more difficult as the complexity of the problem rises, indicated in a pithy analogy:
Solving optimization problems is like being left blindfolded at the top of a hill and asked to walk to the ocean. One way to do that is to judge where to go by feeling around with your foot and taking one step a time in the steepest downward direction. That works if there’s only one hill. But if you’re at the top of the Himalayas, you’ll quickly get stuck because there are thousands of peaks and foothills and flat parts. That’s essentially what happens with current optimization algorithms.
“If you’re lucky, maybe you’ll wind up in the sea but more likely you’ll wind up in a valley or a lake,” said Domingos. “If you could see the whole landscape, you’d say ‘oh, here’s where I have to go,’ but the problem is you can’t see everywhere and neither can today’s algorithms.”
This analogy will sound familiar to those acquainted with evolutionary biology. The “fitness landscape” metaphor coined by Sewall Wright deals with hills and valleys, too; in that context, peaks represent high fitness and valleys represent low fitness. Wright and subsequent evolutionary theorists realized that organisms could get stuck on local fitness peaks (optima). Reaching a higher fitness peak would require losing fitness temporarily to cross a valley (see “Complexity by Subtraction“). And yet it appears frequently that animals did cross valleys, losing and regaining traits multiple times according to phylogenetic trees. Could organisms reach high levels of fitness by blind, unguided processes?
It’s obvious that Friesen and Domingos did not design RDIS to rely on blind search. It was intelligently designed, in multiple ways, throughout. First, they had to know a lot about existing optimization tools. Second, they had to understand their goal. Third, they had to design multiple interacting factors to achieve their goal. It’s also a case of hierarchical design. In a real sense, they had to optimize their optimization tool.
Intelligent designers know how to cross valleys to reach higher peaks, because they can see a distant goal and aim for it. Natural selection has no foresight. Portraying neo-Darwinism’s mechanisms as optimization strategies is, therefore, wishful thinking (see “Evolutionary Computing: The Invisible Hand of Intelligence“). Dembski has shown in Being as Communion and in his earlier book No Free Lunch that no search strategy is superior to blind search — unless information is added to the system by intelligence.
Optimization is a good example of intelligent design at work in the sciences. Not only is the field permeated with ID concepts from start to finish, it is also highly useful and widely applicable. So is ID a “science stopper” as Darwinists often claim? Quite to the contrary.
Image: New York Stock Exchange, by Ryan Lawler (Own work) [Public domain], via Wikimedia Commons.