Criteria of machine learning applied to animal learning
The machine learning literature provides concrete criteria for identifying if an algorithm has learned (Thrun & Pratt 1998). Specifically, given (1) a task , (2) training experience , and (3) a performance measure , if performance at the task improves with experience, the algorithm is said to have learned. This is a useful framework for interpreting observational animal movement data. For example, for the sheep and moose in Jesmer et al. (2018) thetask was maximizing energy intake and the training experience was several years of moving around the landscape. Theperformance measure was the question: Did the animals adopt a migratory movement strategy to track variability in energy availability across space and time? Because of an increase in the proportion of migrants in the population over time (and, thereby, an increase in the proportion of individuals with increased energy intake), the animals likely had “learned”.
A major challenge to applying machine learning criteria to moving animals involves identifying the task and performance measure in meaningful ways, given the animals’ spatial context and scale of movement. Survival and reproduction are the ultimate tasks, but foraging, resting, finding a mate, and avoiding predation are all proximal tasks. Nonetheless, the framework helpfully and unambiguously associates movement in the environment with training experience . Table 2 cross-references a machine-learning example with field studies that provided experimental evidence of learning.
Cognitive ecologists typically have more stringent criteria for identifying learning. For example, experimentation plus control conditions sufficient to rule out alternate explanations are fundamental to confirming the existence of social learning (Reader & Biro 2010). In this framework, experimentation could involve manipulation of physical aspects of the environment, individual animals via translocations or similar means, or the routes governing social transmission of information.
A particularly difficult challenge involves applying the statistical learning criteria to problems where the learning involves updating the world model (as described above) in a familiar landscape rather than learning about a fundamentally novel landscape. For example, in the foraging models of Bracis et al. (2015, 2018), the task is maximization of instantaneous energy intake, the trainingexperience is the movement (together with the acquisition of information for updating the cognitive map), and the performance measure is the amount of forage obtained. However, because the modeled movement process is stationary (in the stochastic sense, meaning that the underlying movement parameters do not change over time), there is no measurable improvement over the long term, merely a constant updating. Nevertheless, without learning (i.e., without the ability to update the cognitive map), the forager performs much worse. Exploring these theoretical discoveries in field systems appears quite challenging because it is difficult to experimentally manipulate the map updating. This contrasts with the heightened feasibility of experimental studies in which learning can be assessed as improvement in animals’ performance as a function of time (see below; Fig. 3).