Human search in a fitness landscape: How to assess the difficulty of a search problem

Computational modeling is widely used to study how humans and organizations search and solve problems in fields such as economics, management, cultural evolution, and computer science. We argue that current computational modeling research on human problem-solving needs to address several fundamental issues in order to generate more meaningful and falsifiable contributions. Based on comparative simulations and a new type of visualization of how to assess the nature of the fitness landscape, we address two key assumptions that approaches such as the NK framework rely on: that the NK captures the continuum of the complexity of empirical fitness landscapes, and that search behavior is a distinct component, independent from the topology of the fitness landscape. We show the limitations of the most common approach to conceptualize how complex, or rugged, a landscape is, as well as how the nature of the fitness landscape is fundamentally intertwined with search behavior. Finally, we outline broader implications for how to simulate problem-solving.

💡 Research Summary

This paper revisits the use of fitness‑landscape models for understanding how humans and organizations search for solutions to complex problems. While the NK framework has become a standard tool across economics, management, cultural evolution, and computer science, the authors argue that it rests on two problematic assumptions: (1) that varying the NK parameters (N and K) adequately captures the full continuum of empirical landscape complexity, and (2) that search behavior can be treated as an independent component separate from the topology of the landscape. To test these assumptions, the authors conduct comparative simulations using both synthetic NK landscapes and empirically derived landscapes built from real human‑decision data (e.g., puzzle solving, investment choices). They also introduce a novel visualization technique that maps fitness values onto a two‑dimensional grid with color gradients, contour lines, and slope indicators, allowing researchers to see at a glance where peaks, valleys, and “traps” lie and how steep the surrounding terrain is. This visual approach reveals that landscapes with the same K value can differ dramatically in peak distribution, basin depth, and connectivity, features that traditional ruggedness metrics (autocorrelation, number of local optima) fail to capture.

The simulation experiments explore a suite of search strategies: random walk, local hill‑climbing, meta‑heuristics such as simulated annealing, and memory‑based approaches that reuse previously successful moves. Results show that the same landscape can yield vastly different outcomes depending on the strategy employed. For instance, meta‑heuristics can escape deep traps in highly rugged terrains, whereas memory‑based agents often converge prematurely on suboptimal peaks. Conversely, even modest changes in the interaction structure of an NK landscape (while keeping K constant) produce large variations in the number and location of local optima, directly affecting the success rates of all strategies.

From these observations the authors construct a “Search Difficulty Index” (SDI). The SDI combines structural properties of the landscape (peak count, average gradient, trap depth) with characteristics of the search algorithm (exploration breadth, use of memory, stochasticity). Empirically, higher SDI values correlate with lower solution quality, longer convergence times, and greater sensitivity to the choice of algorithm—relationships that cannot be explained by NK parameters alone.

In the discussion, the paper emphasizes that treating the landscape and the search process as separable entities is untenable. Instead, they are tightly coupled: the topology shapes which strategies are viable, and the strategies in turn determine which parts of the topology are explored and how the perceived difficulty evolves. The new visualization and SDI provide practical tools for researchers to pre‑characterize a problem space, select appropriate algorithms, and design experiments that are more falsifiable.

The conclusion outlines future research directions: expanding the empirical dataset to cover a broader range of domains, standardizing the visualization and SDI methodology, modeling dynamic landscapes that evolve as agents interact, and investigating co‑evolution of search strategies. By reframing human problem‑solving as an interaction between search behavior and landscape structure, the authors argue that more realistic, testable, and policy‑relevant models can be built, moving the field beyond the limited explanatory power of the classic NK framework.