Statistical Learning: Machine Learning Approaches and Beyond
Statistical learning (Hastie et al . 2009) is a branch of machine learning concerned with the development and study of algorithms to perform specific tasks with minimal instruction. The tasks involve an explicit goal, such as parameter estimation or classification, and require a clear objective function, such as minimizing a cost function or correctly classifying data. To the extent that animals also have clear objective functions (e.g., ultimately: increasing individual fitness; proximally: eating, avoiding being eaten, reproducing), and that these objectives might be satisfied by performing a specific movement-related task (e.g., selecting appropriate places to forage), it is useful to draw a general analogy between a machine-learning algorithm and an animal that learns. As described above, we use the termtask-based learning when referring to this type of process.
A wide range of machine learning approaches emphasize the importance of improvement through experience (Jordan & Mitchell 2015), which is close to some definitions of animal learning. Good examples are artificial neural networks (ANN), a class of biologically inspired learning algorithms. The input of an ANN, typically the sensory perception of the agent or animal, is propagated through a network of idealized neurons, which is readjusted by experience-generated reward signals. The output of the ANN induces observable behaviour.
Another learning-like algorithm is a Bayesian probabilistic model for inference (which can, incidentally, also drive an ANN). While Bayesian reasoning is most often applied for statistical tasks such as parameter estimation and complex model fitting, it is also viewed as a central, probabilistic model for human cognition and learning (Chater et al. 2006; Tenenbaum et al. 2006). In the specific context of animal movement, prior information represents existing knowledge or existing preference sets (e.g., spatial memory and selection coefficients). Bayesian perspectives readily permit such prior knowledge to be updated with new data (experiences) gained by an animal’s movement through the environment. For example, Michelot et al. (2019) draw an explicit analogy between stochastic rule-based animal movement and a Gibbs sampler performing Markov chain Monte Carlo sampling. The resulting posterior distributions accurately reflect the animal’s resource selection function (RSF). The equivalence between an optimizing algorithm and an animal gaining familiarity with its landscape provides an interesting template. One could generate a similar animal (sampler) that does update its movement coefficients based on the mismatch between its experiences and the environment. Box 3 builds on these rule-based decision-making ideas to draw connections between mobile autonomous robots and learning animals.