Explaining Exploration–Exploitation in Humans

• Published in: Big data and cognitive computing Vol. 6; no. 4; p. 155
• Main Authors: Candelieri, Antonio, Ponti, Andrea, Archetti, Francesco
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.12.2022
• Subjects: active human learning; Approximation; Budgets; Decision analysis; Decision making; Decision trees; Expected utility; explainable machine learning; Exploitation; exploration–exploitation dilemma; Global optimization; Human acts; Human behavior; Machine learning; Optimization; Pareto efficiency; Pareto optimum; Searches and seizures; Uncertainty
• ISSN: 2504-2289; 2504-2289
• DOI: 10.3390/bdcc6040155

Human as well as algorithmic searches are performed to balance exploration and exploitation. The search task in this paper is the global optimization of a 2D multimodal function, unknown to the searcher. Thus, the task presents the following features: (i) uncertainty (i.e., information about the function can be acquired only through function observations), (ii) sequentiality (i.e., the choice of the next point to observe depends on the previous ones), and (iii) limited budget (i.e., a maximum number of sequential choices allowed to the players). The data about human behavior are gathered through a gaming app whose screen represents all the possible locations the player can click on. The associated value of the unknown function is shown to the player. Experimental data are gathered from 39 subjects playing 10 different tasks each. Decisions are analyzed in a Pareto optimality setting—improvement vs. uncertainty. The experimental results show that the most significant deviations from the Pareto rationality are associated with a behavior named “exasperated exploration”, close to random search. This behavior shows a statistically significant association with stressful situations occurring when, according to their current belief, the human feels there are no chances to improve over the best value observed so far, while the remaining budget is running out. To classify between Pareto and Not-Pareto decisions, an explainable/interpretable Machine Learning model based on Decision Tree learning is developed. The resulting model is used to implement a synthetic human searcher/optimizer successively compared against Bayesian Optimization. On half of the test problems, the synthetic human results as more effective and efficient.