‘Control’ means different things in different data science communities
In his recent book Human Compatible, leading AI researcher and computer scientist Stuart Russell describes the dangers of “overly intelligent” algorithms, and proposes a solution based on incorporating uncertainty into the algorithms’ “understanding” of human preferences. Russell is puzzled how statisticians, control theory researchers, and operations researchers haven’t thought of this:
“In all the work on utility maximization, and loss function, the reward function, and the loss function are known perfectly. How could this be? How could the AI community (and the control theory, operations research, and statistics communities) have such a huge blind spot for so long, even while embracing uncertainty in all other aspects of decision making?” (p. 176)
In my recent work ‘Improving’ prediction of human behavior using behavior modification I stumbled upon a similar arid land in trying to use statistical notation to describe the combination of two operations used by digital platform: prediction and modification. The lack of language puzzled me. Furthermore, my work has received surprised reactions by statisticians.
Why the surprise?
The answer has to do with the term ‘control’ — what it means to the different communities, and the role of control in algorithms/models.
Let me share my insights about the meaning of ‘control’ to different research communities, from my journey as a trained statistician, collaborating with researchers in social science, human-computer interaction, and machine learning, while also keeping an eye on what operations research and industrial engineering colleagues are talking about.
For statisticians, ‘control’ is typically a feature related to the researcher’s ability to control the study design. For example, in a randomized experiment, the researcher can exert control by choosing a mechanism for assigning subjects to treatments. In drawing a sample from a population, a researcher might use stratified sampling (e.g. by gender), which gives them more control over the resulting sample breakdown. ‘Control’ is therefore related to the researcher’s ability to assert control over the study design, typically to improve the ability of the resulting data to answer the question of interest.
Another use of ‘control’ is by industrial statisticians: statistical quality control or statistical process control (SPC) focuses on monitoring manufacturing processes for the purpose of detecting changes in the underlying behavior that lead to unacceptable product quality. “Out of control” relates to the manufacturing process generating the products. Here control is a feature of the production environment. Industrial statisticians also try to ‘control’ the escape of unacceptable products by employing acceptance sampling schemes (such as the popular ISO 2859–1) for inspecting samples of product batches. Here control is by either the manufacturer who controls the out-flow of unacceptable product batches, or by the customer, who controls the in-flow of unacceptable batches.
In social science, the term ‘control’ is most commonly used in “control variables” — these are additional measurements on subjects, collected and used for the purpose of “controlling” against confounding in estimating/testing causal effects from observational data. Control here is associated with the researcher’s intention to statistically ‘control’ for confounding variables.
Note that in none of the above uses is control used in an interactive way, with feedback loops between the controller and controlled.
In Human-Computer Interaction (HCI), a ‘control’ can refer to a device a user utilizes for interacting with a system (e.g. a mouse for controlling the cursor). Control is therefore attributed to the user for whom the interface is intended.
Turning back from humans to ‘things’, in operations research / industrial engineering, control theory “deals with the control of dynamical systems in engineered processes and machines.” Here we have moved to an interaction (“dynamic”) between a controlled device, and a “controller” that tries to keep the device’s behavior within some range. Two familiar examples of controllers are a thermostat regulating the room temperature and cruise control that control’s a car’s speed.
Finally, we reach the machine learning meaning of ‘control’, specifically, in the context of reinforcement learning. Based on the idea from control theory, one main difference is replacing the mathematical control models with data-driven predictive model (most recently using deep learning). A critical second difference is the application of this form of control to humans rather than devices. We now have feedback loops between machine learning algorithms that predict user behavior, and behavior modification techniques implemented via software, that “regulate” (i.e. push) users’ behavior towards a per-specified objective function.
The Tower of Babel is now exposed: except for machine learners and AI researchers using ‘control’ to mean reinforcement learning-type control, no other research community even considers predicting and manipulating human behavior in an interactive way. That is also why there is insufficient language to represent such operations and combinations. I hope this post serves as a first step towards breaking the data science disciplinary boundaries, and awakening all the non-reinforcement-learning researchers to the critical role of RL systems in our lives and future.