The U.S. Defense Advanced Research Projects Agency (DARPA) has launched a program that will create the technology to make new generations of artificial intelligence (AI) systems "explainable." DARPA'S Explainable AI (XAI) program aims to create new machine learning methods to produce more explainable models and combine them with explanation techniques. And why the need to understand AI? That's because explainable AI -- especially explainable machine learning -- will be essential if future American warfighters are to understand, appropriately trust and effectively manage an emerging generation of AI "partners" such as battlefield robots and machines. XAI is vital because continued advances in AI promise to produce autonomous systems that will perceive, learn, decide and act on their own. The effectiveness of these AI systems, however, is limited by the machine's current inability to explain their decisions and actions to human users.

In a normative environment an agent's actions are not only directed by its goals but also by the norms imposed on the agent. However, the potential conflicts within and between the agent's goals and norms makes decision-making in these frameworks a challenging task. The questions we are addressing in this paper are: (i) how should an agent act in a normative environment? We propose a solution in which a normative planning problem serves as the basis for a practical reasoning approach based on argumentation. The properties of the best plan(s) with respect to goal achievement and norm compliance are mapped to arguments that are used to explain why a plan is justified, using an existing proof dialogue game.

This paper studies two issues concerning relevance in structured argumentation in the context of the ASPIC framework, arising from the combined use of strict and defeasible inference rules. One issue arises if the strict inference rules correspond to classical logic. A longstanding problem is how the trivialising effect of the classical Ex Falso principle can be avoided while satisfying consistency and closure postulates. In this paper, this problem is solved by disallowing chaining of strict rules, resulting in a variant of the ASPIC framework called ASPIC*, and then disallowing the application of strict rules to inconsistent sets of formulas. Another issue is minimality of arguments.

We propose a general model explanation system (MES) for "explaining" the output of black box classifiers. This paper describes extensions to Turner (2015), which is referred to frequently in the text. We use the motivating example of a classifier trained to detect fraud in a credit card transaction history. The key aspect is that we provide explanations applicable to a single prediction, rather than provide an interpretable set of parameters. We focus on explaining positive predictions (alerts). However, the presented methodology is symmetrically applicable to negative predictions.

This paper studies two issues concerning relevance in structured argumentation in the context of the ASPIC+ framework, arising from the combined use of strict and defeasible inference rules. One issue arises if the strict inference rules correspond to classical logic. A longstanding problem is how the trivialising effect of the classical Ex Falso principle can be avoided while satisfying consistency and closure postulates. In this paper, this problem is solved by disallowing chaining of strict rules, resulting in a variant of the ASPIC+ framework called ASPIC*, and then disallowing the application of strict rules to inconsistent sets of formulas. Thus in effect Rescher & Manor's paraconsistent notion of weak consequence is embedded in ASPIC*. Another issue is minimality of arguments. If arguments can apply defeasible inference rules, then they cannot be required to have subset-minimal premises, since defeasible rules based on more information may well make an argument stronger. In this paper instead minimality is required of applications of strict rules throughout an argument. It is shown that under some plausible assumptions this does not affect the set of conclusions. In addition, circular arguments are in the new ASPIC* framework excluded in a way that satisfies closure and consistency postulates and that generates finitary argumentation frameworks if the knowledge base and set of defeasible rules are finite. For the latter result the exclusion of chaining of strict rules is essential. Finally, the combined results of this paper are shown to be a proper extension of classical-logic argumentation with preferences and defeasible rules.

Adding weights or preferences to Abstract Argumentation Frameworks can help disentangle semantics from otherwise all-equivalent attacks. Having such information makes possible to distil the set of found extensions by reducing their number. In this work we provide a new definition of weighted defence: according to it, all the attacks from an argument to a set of arguments are considered with a single global weight, i.e., attacks are grouped together. This provides a coherent view w.r.t. defence, which is usually “collective” in the literature. Moreover, we model weighted defences from related works in the same algebraic framework: this allows us to compare all the different proposals together.

We present a novel approach to account for preferences in a well known structured argumentation formalism, Assumption-Based Argumentation (ABA). The new formalism, called ABA+, incorporates object-level preferences (over assumptions) directly into the attack relation to reverse attacks. We give several basic desirable properties of ABA+.

In this report from the field we describe jArgSemSAT, a Java re-implementation of ArgSemSAT. We show that jArgSemSAT can be easily integrated in existing argumentation systems (1) as an off-the-shelf, standalone, library; (2) as a Tweety compatible library; and (3) as a fast and robust web service freely available on the Web. The performance section shows that — despite being written in Java — jArgSemSAT is very efficient w.r.t. preferred semantics, which has associated problems with high computational complexity.

In the study of argumentation-based reasoning, argument justification has received far more attention than statement justification, often treated as a simple byproduct of the former. As a consequence, counterintuitive results and significant losses of sensitivity can be identified in the treatment of statement justification by otherwise appealing formalisms. To overcome this limitation, we propose to reappraise statement justification as a formalism-independent component. To this purpose, we introduce a novel general model of argumentation-based reasoning based on multiple levels of labellings, one of which is devoted to statement justification. This model is able to encompass several literature proposals as special cases: we illustrate this ability for the case of the ASPIC+ formalism and provide a first example of tunable statement justification in this context.

Reasoning with the maximally consistent subsets (MCS) of the premises is awell-known approach for handling contradictory information. We introduce two argumentation-based methods for doing so: a declarative approach that is related to Dung-style semantics for abstract argumentation, and a computational approach that is based on extensions of Gentzen-type proofs systems. This brings about a new perspective on reasoning with MCS which shows a strong link between the latter and argumentation systems, and which can be extended to related formalisms. A by-product of this is the introduction of a dynamic proof system for classical logic and rebuttal attacks, which is sound and complete with respect to Dung's stable semantics for the associated argumentation framework.