Trust and trustworthy artificial intelligence: A research agenda for AI in the environmental sciences.

Demands to manage the risks of artificial intelligence (AI) are growing. These demands and the government standards arising from them both call for trustworthy AI. In response, we adopt a convergent approach to review, evaluate, and synthesize research on the trust and trustworthiness of AI in the environmental sciences and propose a research agenda. Evidential and conceptual histories of research on trust and trustworthiness reveal persisting ambiguities and measurement shortcomings related to inconsistent attention to the contextual and social dependencies and dynamics of trust. Potentially underappreciated in the development of trustworthy AI for environmental sciences is the importance of engaging AI users and other stakeholders, which human-AI teaming perspectives on AI development similarly underscore. Co-development strategies may also help reconcile efforts to develop performance-based trustworthiness standards with dynamic and contextual notions of trust. We illustrate the importance of these themes with applied examples and show how insights from research on trust and the communication of risk and uncertainty can help advance the understanding of trust and trustworthiness of AI in the environmental sciences.

On the existence of momentum in professional football.

Using United States National Football League play-by-play data for the 2002-2012 seasons, we train a neural network to predict win probability, based on measures of the game state. This predictor's performance is comparable to the point spread at the start of the game and improves thereafter with little bias. We define a measure of success as the change in a team's win probability over the course of a possession, and show that streaks in this measure are highly unlikely to be random. Further, this finding holds when controlling for the effects of clock management in the fourth quarter of football games, when win probability can increase incrementally for the leading team as the game continues. By defining momentum as an increase in win probability over the course of at least three successive changes in possession, we show some ability to anticipate its emergence, based on game state, using a second neural network. The possibility of using this knowledge for strategic advantage is discussed. We consider these results in the context of examples from National Football League games, including that from Super Bowl LI (Atlanta Falcons versus the New England Patriots), and end with some discussion of future extensions to this work.

Does the principle of investment diversification apply to the starting pitching staffs of major league baseball teams?

Financial advisors often emphasize asset diversification as a means of limiting losses from investments that perform unexpectedly poorly over a particular time period. One might expect that this perceived wisdom could apply in another high stakes arena-professional baseball-where player salaries comprise a substantial portion of a team's operational costs, year-to-year player performance is highly variable, and injuries can occur at any time. These attributes are particularly true in the case of the starting pitching staffs of professional baseball teams. Accordingly, this study analyzes starting pitcher performance and financial data from all Major League Baseball teams for the period 1985-2016 to determine whether the standard investment advice is applicable in this context, understanding that the time horizon for success for an investor and a baseball team may be distinct. A multiple logistic regression model of playoff qualification probability, based on realized pitcher performance, measures of luck, and starting pitcher staff salary diversification is used to address this question. A further stratification is conducted to determine whether there are differences in strategy for teams with allocated financial resources that are above or below league average. We find that teams with above average resources increase their post-season qualification probability by focusing their salary funds on a relative few starting pitchers rather than diversifying that investment across the staff. Second, we find that pitcher performance must align with that investment in order for the team to have a high qualification probability. Third, the influence of luck is not negligible, but those teams that allocate more overall funds to their pitching are more resilient to bad luck. Thus, poorly resourced teams, who are generally unable to bid for pitchers at the highest salary levels, must adopt alternative strategies to maintain their competitiveness.

NBA team home advantage: Identifying key factors using an artificial neural network.

What determines a team's home advantage, and why does it change with time? Is it something about the rowdiness of the hometown crowd? Is it something about the location of the team? Or is it something about the team itself, the quality of the team or the styles it may or may not play? To answer these questions, season performance statistics were downloaded for all NBA teams across 32 seasons (83-84 to 17-18). Data were also obtained for other potential influences identified in the literature including: stadium attendance, altitude, and team market size. Using an artificial neural network, a team's home advantage was diagnosed using team performance statistics only. Attendance, altitude, and market size were unsuccessful at improving this diagnosis. The style of play is a key factor in the home advantage. Teams that make more two point and free-throw shots see larger advantages at home. Given the rise in three-point shooting in recent years, this finding partially explains the gradual decline in home advantage observed across the league over time.

A virtual climate library of surface temperature over North America for 1979-2015.

The most comprehensive continuous-coverage modern climatic data sets, known as reanalyses, come from combining state-of-the-art numerical weather prediction (NWP) models with diverse available observations. These reanalysis products estimate the path of climate evolution that actually happened, and their use in a probabilistic context-for example, to document trends in extreme events in response to climate change-is, therefore, limited. Free runs of NWP models without data assimilation can in principle be used for the latter purpose, but such simulations are computationally expensive and are prone to systematic biases. Here we produce a high-resolution, 100-member ensemble simulation of surface atmospheric temperature over North America for the 1979-2015 period using a comprehensive spatially extended non-stationary statistical model derived from the data based on the North American Regional Reanalysis. The surrogate climate realizations generated by this model are independent from, yet nearly statistically congruent with reality. This data set provides unique opportunities for the analysis of weather-related risk, with applications in agriculture, energy development, and protection of human life.

Peer review, program officers and science funding.

Increased competition for research funding has led to growth in proposal submissions and lower funding-success rates. An agent-based model of the funding cycle, accounting for variations in program officer and reviewer behaviors, for a range of funding rates, is used to assess the efficiency of different proposal-submission strategies. Program officers who use more reviewers and require consensus can improve the chances of scientists submitting fewer proposals. Selfish or negligent reviewers reduce the effectiveness of submitting more proposals, but have less influence as available funding declines. Policies designed to decrease proposal submissions reduce reviewer workload, but can lower the quality of funded proposals. When available funding falls below 10-15% in this model, the most effective strategy for scientists to maintain funding is to submit many proposals.