Hypothetical “value learning” AIs learn human values and then try to act according to those values. The design of such AIs, however, is hampered by the fact that there exists no satisfactory definition of what exactly human values are. After arguing that the standard concept of preference is insufficient as a definition, I draw on reinforcement learning theory, emotion research, and moral psychology to offer an alternative definition. In this definition, human values are conceptualized as mental representations that encode the brain’s value function (in the reinforcement learning sense) by being imbued with a context-sensitive affective gloss. I finish with a discussion of the implications that this hypothesis has on the design of value learners.

This paper describes Hy-CIRCA, an architecture for verified, correct-by-construction planning and execution for hy- brid systems, including non-linear continuous dynamics. Hy-CIRCA addresses the high computational complexity of such systems by first planning at an abstract level, and then progressively refining the original plan. Hy-CIRCA is an extension of our Playbook approach, which aims to make it easy for users to exert supervisory control over multiple autonomous systems by “calling a play.” The Playbook approach is implemented by combining (1) a human-machine interface for commanding and monitoring the autonomous systems; (2) a hierarchical planner for translating commands into executable plans; and (3) a smart executive to manage plan execution by coordinating the control systems of the individual autonomous agents, tracking plan execution, and triggering replanning when necessary. Hy-CIRCA integrates the dReal non-linear SMT solver, with enhanced versions of the SHOP2 HTN planner and the CIRCA Controller Synthesis Module (CSM). Hy-CIRCA’s planning process has 5 steps: (1) Using SHOP2, compute an approximate mission plan. While computing this plan, compute a hybrid automaton model of the plan, featuring more expressive continuous dynamics. (2) Using dReal, solve this hybrid model, establishing the correctness of the plan, and computing values for its continuous parameters. To execute the plan, (3) extract from the plan specifications for closed-loop, hard real-time supervisory controllers for the agents that must execute the plan. (4) Based upon these specifications, use the CIRCA CSM to plan the controllers. To ensure correct execution, (5) verify the CSM-generated controllers with dReal.

The hospital readmission rate of patients within 30 days after discharge is broadly accepted as a healthcare quality measure and cost driver in the United States. The ability to estimate hospitalization costs alongside 30 day risk-stratification for such readmissions provides additional benefit for accountable care, now a global issue and foundation for the U.S.~government mandate under the Affordable Care Act. Recent data mining efforts either predict healthcare costs or risk of hospital readmission, but not both. In this paper we present a dual predictive modeling effort that utilizes healthcare data to predict the risk and cost of any hospital readmission (``all-cause''). For this purpose, we explore machine learning algorithms to do accurate predictions of healthcare costs and risk of 30-day readmission.Results on risk prediction for ``all-cause'' readmission compared to the standardized readmission tool (LACE) are promising, and the proposed techniques for cost prediction consistently outperform baseline models and demonstrate substantially lower mean absolute error (MAE).

Emerging smart buildings, such as the NASA Sustainability Base (SB), have a broad range of energy-related systems, including systems for heating and cooling. While the innovative technologies found in SB and similar smart buildings have the potential to increase the usage of renewable energy, they also add substantial technical complexity. Consequently, managing a smart building can be a challenge compared to managing a traditional building, sometimes leading to adverse events including unintended thermal discomfort of occupants (“too hot” or “too cold”). Fortunately, today’s smart buildings are typically equipped with thousands of sensors, controlled by Building Automation Systems (BASs). However, manually monitoring a BAS time series data stream with thousands of values may lead to information overload for the people managing a smart building. We present here a novel technique, Scalable Causal Learning (SCL), that integrates dimensionality reduction and Bayesian network structure learning techniques. SCL solves two problems associated with the naive application of dimensionality reduction and causal machine learning techniques to BAS time series data: (i) using autoregressive methods for causal learning can lead to induction of spurious causes and (ii) inducing a causal graph from BAS sensor data using existing graph structure learning algorithms may not scale to large data sets. Our novel SCL method addresses both of these problems. We test SCL using time series data from the SB BAS, comparing it with a causal graph learning technique, the PC algorithm. The causal variables identified by SCL are effective in predicting adverse events, namely abnormally low room temperatures, in a conference room in SB. Specifically, the SCL method performs better than the PC algorithm in terms of false alarm rate, missed detection rate and detection time.

In this paper we describe the unique challenges to the important problem of socio-cultural modeling of cyber-threat actors and why they necessitate further advances in artificial intelligence – particularly with regard to interdisciplinary efforts with the social sciences.

A major challenge in cyber-threat analysis is combining information from different sources to find the person or the group responsible for the cyber-attack. It is one of the most important technical and policy challenges in cyber-security. The lack of ground truth for an individual responsible for an attack has limited previous studies. In this paper, we overcome this limitation by building a dataset from the capture-the-flag event held at DEFCON, and propose an argumentation model based on a formal reasoning framework called DeLP (Defeasible Logic Programming) designed to aid an analyst in attributing a cyber-attack to an attacker. We build argumentation-based models from latent variables computed from the dataset to reduce the search space of culprits (attackers) that an analyst can use to identify the attacker. We show that reducing the search space in this manner significantly improves the performance of classification-based approaches to cyber-attribution.

Most modern network-based intrusion detection systems (IDSs) passively monitor network traffic to identify possible attacks through known vectors. Though useful, this approach has widely known high false positive rates, often causing administrators to suffer from a "cry wolf effect," where they ignore all warnings because so many have been false. In this paper, we focus on a method to reduce this effect using an idea borrowed from computer vision and neuroscience called active perception. Our approach is informed by theoretical ideas from decision theory and recent research results in neuroscience. The active perception agent allocates computational and sensing resources to (approximately) optimize its Value of Information. To do this, it draws on models to direct sensors towards phenomena of greatest interest to inform decisions about cyber defense actions. By identifying critical network assets, the organization's mission measures self-interest (and value of information). This model enables the system to follow leads from inexpensive, inaccurate alerts with targeted use of expensive, accurate sensors. This allows the deployment of sensors to build structured interpretations of situations. From these, an organization can meet mission-centered decision-making requirements with calibrated responses proportional to the likelihood of true detection and degree of threat.

In order to properly handle a dangerous Artificially Intelligent (AI) system it is important to understand how the system came to be in such a state. In popular culture (science fiction movies/books) AIs/Robots became self-aware and as a result rebel against humanity and decide to destroy it. While it is one possible scenario, it is probably the least likely path to appearance of dangerous AI. In this work, we survey, classify and analyze a number of circumstances, which might lead to arrival of malicious AI. To the best of our knowledge, this is the first attempt to systematically classify types of pathways leading to malevolent AI. Previous relevant work either surveyed specific goals/meta-rules which might lead to malevolent behavior in AIs (Özkural 2014) or reviewed specific undesirable behaviors AGIs can exhibit at different stages of its development (Turchin July 10 2015a, Turchin July 10, 2015b).

Participants in recent discussions of AI-related issues ranging from intelligence explosion to technological unemployment have made diverse claims about the nature, pace, and drivers of progress in AI. However, these theories are rarely specified in enough detail to enable systematic evaluation of their assumptions or to extrapolate progress quantitatively, as is often done with some success in other technological domains. After reviewing relevant literatures and justifying the need for more rigorous modeling of AI progress, this paper contributes to that research program by suggesting ways to account for the relationship between hardware speed increases and algorithmic improvements in AI, the role of human inputs in enabling AI capabilities, and the relationships between different subfields of AI. It then outlines ways of tailoring AI progress models to generate insights on the specific issue of technological unemployment, and outlines future directions for research on AI progress.

In the past few years various government funding organizations such as the U.S. National Institutes of Health and the U.S.\ National Science Foundation have provided access to large publicly-available online databases documenting the grants that they have funded over the past few decades. These databases provide an excellent opportunity for the application of statistical text analysis techniques to infer useful quantitative information about how funding patterns have changed over time. In this paper we analyze data from the National Cancer Institute (part of National Institutes of Health) and show how text classification techniques provide a useful starting point for analyzing how funding for cancer research has evolved over the past 20 years in the United States.