Learning To Climb On The Fitness Landscape

Stephen Heap gives an overview of Acerbi, Tennie and Mesoudi, 2016, "Social learning solves the problem of narrow-peaked search landscapes: experimental evidence in humans".

Nature presents a diversity of challenges that individuals can address through learning, and the best learning strategies to employ can depend on the type of challenge. Challenges can vary in how rewarding a given solution is, with respect to how closely it resembles some optimal ideal. In some cases, small incremental improvements toward the ideal provide adequate feedback between solution and reward for an individual to learn asocially (wide fitness landscape). By contrast, individuals have inadequate guidance for tracking an ideal solution and rely more on social learning when only some solutions are rewarded (narrow fitness landscape).

Acerbi, Tennie and Mesoudi tested this proposition with an experiment on human subjects in a simulated hunt, for which success was dependent on the design of an arrowhead. Subjects were informed of how well their arrowhead performed on a hunt, and could redesign their arrowheads over a hunting season. The experiment employed a two-factor design. Firstly, whilst all individuals were given feedback for their own designs, one condition allowed individuals access to social information on the arrowhead designs of others. The demonstrators in this case were actually fabricated as part of the experimental design in order to control for the content of social information. Secondly, the scoring rule that relates an arrowhead design to its success in hunting produced one of two types of fitness landscape. Wide fitness landscapes rewarded a range of solutions, whilst rewards were restricted to very specific solutions when the fitness landscape was narrow.

Tests reveal that individual learning performs poorly in narrow-peaked landscapes, yet social learning can act as a kind of buffer that protects against getting stuck with a poor solution. That is, whilst social learning can be advantageous in a range of conditions, its advantages can become more pronounced as the fitness landscape narrows. The paper advocates that we pay greater attention to the fitness landscape for a given challenge when wanting to understand the evolution of social learning. Different landscapes are likely to give rise to different sets of learning strategies, and the nature of challenges being faced by early humans may have influenced the shape of the evolving cognitive toolkit.

REFERENCES

Acerbi, A., Tennie, C., & Mesoudi, A. (2016). Social learning solves the problem of narrow-peaked search landscapes: experimental evidence in humans. Royal Society Open Science, 3(9), 160215.