Quantifying effects of tasks on group performance in social learning

The study investigates how characteristics of learning tasks influence group performance under social learning, especially when social learning is modeled as imperfect imitation rather than perfect copying. Prior research has emphasized social learning strategies and network structures but often overlooked task features and assumed perfect imitation. This work addresses three questions: (1) How do task features (type, complexity, granularity) affect group performance? (2) What performance levels can groups achieve via social learning? (3) How can the effects of different variables on group performance be quantified? To answer these, the authors construct a multi-agent model where individuals with status quo bias interact on networks and search rugged landscapes representing two task types: hard-to-easy and easy-to-hard. The importance of variables is quantified via decision trees, aiming to clarify the decisive role of tasks in cultural evolution.

Prior work shows social learning as central to cultural evolution and has explored social learning strategies (e.g., best, conformity) and network structure effects on group performance. Many studies used the NK model to generate rugged landscapes but assumed perfect imitation and focused on complexity while often neglecting other task features. Some experiments examined single task types or treated differences as only complexity. Recent findings suggest social learning often produces novel solutions by combining information from multiple sources, implying imperfect imitation is prevalent. Interactions among variables (strategies, networks) complicate disentangling effects. The authors argue task granularity has been overlooked in social learning, despite cognitive science indicating humans operate on discrete representations with perceptual limits, affecting how tasks are perceived and decomposed.

Agent-based model of social learning on rugged landscapes. Task environments: Use transformed one-dimensional optimization test functions to create deterministic multi-peak landscapes—Ackley (hard-to-easy) and Rastrigin (easy-to-hard). Both are rescaled to solution/payoff in [0,1] with a single global optimum. Task complexity C in {1,...,14} maps to number of peaks p = 2^C − 1 (from 1 to 16383), categorized as low (1–5), medium (6–10), high (11–14). Task granularity ω ∈ {1,...,10} partitions [0,1] into 10^ω equally spaced discrete candidate solutions; higher ω implies finer granularity and more solutions. Networks: Four network types considered—fully connected (FC), locally connected lattice, Watts–Strogatz (WS), Barabási–Albert (BA). Main text uses FC with N=100; supplementary analyses vary N (50–5000), lattice average degree (10–98), and WS rewiring probability (0–1), keeping average degree K=10 where applicable (FC has K=N−1; WS rewiring=0.1 in baseline comparisons). Group conservativeness (status quo bias, SQB): Each individual i has φ_i ∈ [0,1]. Larger φ indicates greater preference for current solution (more conservative). Authors define conservative as φ<0.5 and open as φ>0.5 (ultra-conservative φ=1 never imitates; ultra-open φ=0 copies exactly). Group-level conservativeness is controlled via distributions over φ: normal (balanced/neutral), left-skewed (conservative-biased), and right-skewed (open-biased) using skew normal distributions. The proportion of conservative individuals varies from 10% to 90%. Social learning mechanism: At each discrete step, an individual samples m neighbor solutions (m=3 in baseline). It chooses a model via one strategy: best (highest payoff), conformity (most frequent), or random. If the chosen neighbor’s payoff exceeds the focal’s, the focal updates by imperfect imitation: s_i(t+1) = φ_i s_i(t) + (1−φ_i) s_j(t), a convex combination of its own solution and the socially acquired solution; otherwise it retains s_i(t). Individuals begin with random initial solutions (payoffs drawn from the task landscape). The process runs for 500 steps per simulation; group performance is the average payoff at each step. Unless stated otherwise, the baseline combination is: FC network, N=100, normal SQB, hard-to-easy task with C=5 and ω=4, best strategy with m=3. Results are averages over 10 repetitions per condition. Decision-tree analysis: To quantify variable effects, two models—CART and Random Forest—predict final group performance y from five inputs: task type T (0 hard-to-easy, 1 easy-to-hard), task granularity ω (1–10), task complexity C (1–14), group conservativeness p (percentage of conservatives 10–90), and strategy S (0 best, 1 conformity, 2 random). One-hot encoding is used for categorical variables. All 7,560 parameter combinations are simulated, each repeated 10 times, yielding 75,600 samples. Data are split 80/20 into train/test. Mean squared error (MSE) guides splits; tree growth stops when rMSE improvement < 0.01, producing a four-layer regression tree. CART and RF yield similar performance and the same tree structure on these data.

• Task type and complexity effects:
  • Hard-to-easy tasks (Ackley): Groups consistently reach the global optimum across complexities; strategies perform similarly well. Complexity does not degrade performance near the optimum due to dominant trend term behavior.
  • Easy-to-hard tasks (Rastrigin): With fine granularity, groups perform well at low complexity (C≤5) but performance declines as complexity rises (C>5) due to many local optima near the global optimum; with coarse granularity, performance fluctuates across complexities.
• Task granularity effects:
  • Across both task types, reducing granularity (coarser tasks) generally lowers performance. In hard-to-easy tasks, performance drops sharply when ω<3 as the solution set becomes sparse and step differences large, inducing learning bottlenecks, especially for conservative individuals.
  • In easy-to-hard tasks of high complexity, coarse-grained tasks can yield slightly better performance than fine-grained tasks because fewer discrete solutions reduce the density of local optima near the global optimum, easing exploration for conservative individuals.
• Group conservativeness:
  • A higher proportion of conservative individuals decreases group performance in both task types. Open individuals reliably reach optimal solutions and outperform conservative individuals.
  • Interaction with granularity: Increasing granularity (higher ω) mitigates conservative individuals’ disadvantages by enabling small, incremental updates that accumulate to high performance, helping conservative-biased groups approach open groups’ outcomes.
• Network-related variables:
  • Network structure, network density, and group size show no significant effects on group performance under the examined conditions (per extensive supplementary simulations).
• Decision-tree quantification of importance:
  • Feature importance from Random Forest: task granularity 36.1%, task complexity 24.8%, task type 17.8% (task features sum to 78.6%); group conservativeness 21.4%; social learning strategy 0.0% (not selected in the regression tree). The first split is on ω at ω>1, indicating strong sensitivity to granularity. Strategy differences are minor overall but are observable in specific cases (e.g., best strategy outperforms others in fine-grained, high-complexity easy-to-hard tasks).

The findings show that properties of the task environment decisively shape group performance under social learning modeled as imperfect imitation. Disentangling task type from complexity reveals that groups excel in hard-to-easy tasks regardless of complexity but struggle in easy-to-hard tasks as complexity rises due to dense local optima near the global optimum. Task granularity, a previously neglected feature, critically governs learnability: finer granularity increases accessible intermediate solutions and feedback, enabling both diffusion and incremental improvement. This especially benefits conservative individuals prone to status quo bias, helping them avoid learning bottlenecks. Decision-tree analyses quantify these relationships, attributing nearly 80% of predictive power to task features and showing group conservativeness as the next most influential factor, while strategy choice has comparatively negligible predictive impact across the full space of conditions. These insights suggest that manipulating task design—particularly increasing granularity—can substantially enhance cultural transmission and group performance, offering practical levers in education, technology adoption, and organizational learning.

This study introduces an agent-based framework of social learning with imperfect imitation and demonstrates that task features—type, complexity, and especially granularity—largely determine achievable group performance. Increasing task granularity improves exploration and mitigates conservative individuals’ bottlenecks, enabling higher collective outcomes. Decision-tree models provide a clear, quantitative decomposition of variable effects, highlighting tasks’ decisive role and the secondary importance of group conservativeness. Future research should integrate individual (asocial) exploration with social learning to assess their relative contributions and examine multi-dimensional task spaces to understand how dimensionality and interdependencies shape performance and diffusion dynamics.

• The model assumes agents rely solely on social learning with no independent (asocial) exploration, potentially underestimating exploration pathways.
• Tasks are one-dimensional; real-world tasks often involve multiple dimensions and interdependencies, which may change landscape structure and learning dynamics.
• Results are based on simulation with specific parameterizations and transformed test functions; empirical validation across domains would strengthen generalizability.
• Strategy importance may be understated in specific real-world contexts not captured by the simulated parameter space.