ISI News
A New Way to Find Better Solutions
A flock of birds can change direction together. Each follows basic rules, and the group moves as one.
Kenneth Zick wondered whether recipes for engineered emergent behavior could be developed, allowing a computer to better solve difficult optimization problems.
The result is Cosm, a computer algorithm, or a set of software instructions, for rapidly finding strong solutions to problems with many possible choices. Cosm gives many simple parts a few novel rules that change over time, then lets their combined behavior guide the search.
Cosm is not generative artificial intelligence. It does not learn from huge collections of data or produce text and images. It repeatedly works on a defined problem.
“Many interacting elements follow simple rules, and through their interactions, something new and unexpected arises,” said Zick, research director of transformational computing at the USC Viterbi School of Engineering and USC Stevens School of Computing and AI’s Information Sciences Institute office in Arlington, Virginia. “It’s more than the sum of its parts.”
The approach could eventually help organizations create a superior strategy for a high-stakes decision or operate power grids more effectively.
How Cosm makes a hard problem navigable
Some problems become difficult because each choice affects many others. Consider partitioning a large set of individuals into two groups, where some pairs must be kept together and some must be kept apart. There may be no way to satisfy all the constraints, and too many possible assignments to evaluate them all.
One of the proposed innovations is separating the constraints into subsets that minimize or avoid conflicting goals, then performing sequences of moves that consider each subset one at a time. Zick compares it to driving with several backseat drivers. If everyone gives directions at once, their conflicting advice may leave the driver unsure where to go. Cosm lets each passenger speak in turn, and, over many rounds, allows an unplanned but effective path to emerge.
A second key ingredient is a way to nudge the search process so that related sets of variables can more freely change values together. Coordinated changes of many variables are difficult to achieve with many existing methods, and can be important to reaching the highest practical solution quality.
Breaking records on difficult tests
Cosm is described in a manuscript titled “Cosm: Orchestrated Collective Dynamics Give Rise to Rapid Computational Search at the Solution Frontier,” written by ISI’s Zick; Professor Nikhil Shukla of the University of Virginia; and Alexander Marakov of Northrop Grumman.
The researchers tested Cosm on difficult optimization benchmarks used to compare problem-solving methods since the late 1990s. Cosm is designed to find high-quality solutions quickly, not prove that no better solution exists. Yet on the largest test, involving 20,000 variables, Cosm found a solution that another program later confirmed was the best possible one. On three other tests, Cosm reached the best previously reported solutions with surprising speed.
Cosm’s biggest breakthrough came on a difficult problem called Gset G64. It has 7,000 variables connected by 41,459 pairwise relationships (edges). The goal is to divide the variables into two groups so the sum of the edge weights between the groups is maximized.
The best score rose from 8,751 in 2015 to 8,752 in 2024. Cosm reached 8,753. Though just one unit higher, an improved score on a problem that researchers have attacked from many angles can represent years of progress.
“When we discovered that new best result for G64, I was jumping up and down,” Zick said.
From laboratory tests to everyday decisions
Because Cosm can solve complex problems on standard computers, users may not need expensive machines built specifically for that task. Finding answers faster could also reduce computing time, electricity use and cost, although researchers have not yet shown broad savings in practical applications.
A business might one day use Cosm to decide which employees should work each shift, which truck should handle each delivery or how a vehicle fleet should travel without wasting time and fuel. A power company might use it to choose where to send repair crews, which could cut costs and reduce the risk of failures.
Much remains to be learned. Cosm has performed best when each choice affects only some others. It has not shown the same advantage when every choice affects nearly everything else. Researchers must learn why it succeeds, which practical problems suit it and whether its laboratory gains hold up in daily use.
After more than 20 years of considering whether collective behavior could be harnessed for computing, Zick sees the findings as a beginning.
“We’re really eager to see how far we can take this,” he said.
Published on November 19th, 2026
Last updated on November 20th, 2026