Exploration and Exploitation in an Artificial Experimenter

An artificial experimenter is a computational implementation of the decision making processes a laboratory experimenter will make. Artificial experimenter’s analyse the available data, propose hypotheses to represent the behaviours investigated and design experiments to evaluate or improve those hypotheses. In doing so they perform active discovery. A key problem faced is deciding when to perform experiments that exploit the information held within the current hypotheses to evaluate them and when to perform experiments that explore the parameter space to discover features of the behaviour being investigated not yet identified. As resources in physical experimentation are extremely limited, addressing this trade-off is critical to obtaining a representative model of the system under investigation. To achieve this, a Bayesian notion of surprise has been used to effectively manage the transition between exploration and exploitation in simulated and physical experimental trials.