Getting the Most Out of Government’s Learning Loop

Over the past 50 years and more, governments have worked to improve decision-making, relying less on anecdote and intuition and more on analysis and evidence. More budgets now focus not just on spending for inputs but also on spending to find and assess the efficiency of alternative ways of reaching desired outcomes, creating the opportunity for true performance measurement. Until the past few years, though, these analytic reforms have progressed slowly, largely because they were expensive and time-consuming.

In this context, recent progress -- especially due to computer-based tools -- has opened new possibilities. The productivity of digital processing has increased some 34 million times since the 1960s, reflecting "Moore's Law," with all but a small fraction of that progress arriving in just the past decade. That digital progress shows no sign of slowing, providing an opportunity for governments to continually improve the measuring of which inputs are producing the best outcomes. Better judgment and decision-making flow from this learning loop, and we clearly need them as technology also brings faster and more disruptive change to jobs and institutions. The tools for enhancing the learning loop are more powerful than ever:

• Better data availability, for both situational awareness and feedback. Not only is vastly more data now stored and available for analysis, but virtually all of the data collected in recent decades is in machine-readable formats.

These changes are amazing. In New York City until recently, for example, it was impossible for snowplow managers or residents to know where the plows were once they moved out of sight; today, GPS systems report exactly what the plows have covered, with maps for the public automatically updated every 15 minutes on the PlowNYC website. Facial and other bio-recognition data are now collected and shared for border and immigration control, health and social-benefits programs, and policing, including sometimes-controversial surveillance applications.

• Better data modelling and expert analysis, to clarify relationships and uncertainties between actions and results. With today's digital tools, we can use data and modelling to measure relevant relationships, including the uncertainties that drive critical risk/reward assessments. We can also reach much further afield for expertise.

Powerful examples are emerging via big data, where regressions and other types of analysis work with much more information. Given enough data, computers can recognize patterns that make it increasingly easy to render speech into written text, translate from one language to another or -- with obvious implications for strategic planning in government - win at chess and Go . Algorithms are getting as good as skilled analysts in interpreting X-rays and searching for relevant medical, legal and administrative information.

• Better data-based decision-making. Already, more decisions are being made via collaborative systems. In other settings, we are using computers to select tax returns for auditing or enforce laws against speeding and running red lights. Computer grading of assignments for the tens of thousands of students who might take an online college course can provide immediate and frequent feedback. And we are well on the way to letting computers operate cars, trucks, mining equipment and drones.

• Better data-based implementation. We can increasingly get better results by focusing on learning during implementation, especially through improved transparency and agility.

Activities are more transparent when more people can see what has happened, thus bringing social influence as well as formal authority to bear. GPS signals from snowplows or other vehicles can make it harder for field crews to slack off. And agility -- the ability to respond rapidly to changing requirements -- is increasingly important as systems require more behavioral change by producers and/or consumers. The FBI's Sentinel Project for electronic case management provided a dramatic case where traditional "waterfall development" (design it all up front) failed dramatically, but agile "small-step-at-a-time" development got the job done.

For the newly powerful learning loop to work well, leaders will need to develop new priorities for data collection and feedback, analytic staffing, engagement of stakeholders in decision-making, and the handing off to algorithms of work now done by humans. The challenges will be difficult. But if we are to succeed in the new digital age, we will clearly need to be able to make choices that are smarter and better-informed than those we can make today. We need to take more systematic and serious advantage of the learning loop.