Paragraph Number

71

In reinforcement learning, data can be generated by an agent's interactions with one or more objects, e.g. a surgeon interacting with a patient. The agent, e.g. the surgeon, can perform an action, and the environment, e.g. a target tissue or a surgical site, can generate one or more observations and, for example, a cost according to some dynamics or parameters, e.g. a tissue removal or an infection risk. The objective can be to discover a treatment, treatment algorithm, treatment modification that can reduce or minimize a measure of the cost, e.g. an infection risk, a patient reported outcome measurement, a functional result. The parameters and dynamics of the environment, e.g. a surgical site, can be unknown, but can be estimated. The environment, e.g. a target tissue or a surgical site, can be modeled as a Markov decision process and actions, with possible probability distributions, e.g. a cost distribution, an observation distribution, and one or more transitions, and a policy or algorithm or solution can be defined as a conditional distribution over actions given one or more observations. Dynamic programming can be coupled with ANNs and applied to multi-dimensional nonlinear problems.

Added by DJM Jan 2024