Explainable Arti cial Intelligence systems, including intelligent agents, must explain their internal decisions to other software agents or to human users. In the latter case, it is necessary that humans understand the outputs returned by the systems. Thus, like it was done in [2] and [3], we use a pseudo-natural language for improving the understanding of the explanations given by software agents. ArgAgent1 is a simulator for Belief-Based Goal Processing (BBGP) [1] which is an extension of the Belief-Desire-Intention (BDI) model [6]. Argumentation is used in the intention formation process, which is comprised of four sequential stages: activation, evaluation, deliberation, and checking. In this demo, we focus on the deliberation stage, whose purpose is to identify con icts among goals and select which goals the agent will commit to. These con icts can be: a) terminal (denoted by t), b) resource (denoted by r), or c) super uity (denoted by s)2. Thus, the aim is to generate explanations for the question: why did the agent commit (or not) to a given goal? Besides, we enrich such explanations with information about the type of con ict that arose between goals. The simulator input is an agent model, composed of initial beliefs, rules, preference values of goals, and plans. Explanatory arguments are constructed based on a set of six-domain-independent deliberation rules , whose premises are uni ed with beliefs that are generated from the Goal Argumentation Framework { which is constructed to determinate the set of chosen goals { and the set of chosen goals, which were obtained after applying a semantics based on con ictfree sets and a function that selects the extension that maximizes utility. The explanations are generated from the set of explanatory arguments and the possible attacks between them, following the method presented in [5]. The pseudo-natural language explanations are built from: (i) the uni ed deliberation rules, which are used to construct explanatory arguments; and (ii) the respective schemes used to generate explanatory sentences from the explanatory arguments.
H. Jasinski et al.
Rule incompatible(g; g0; conflict); preferred(g; g0) ! pursued(g)
<goal name g> and <goal name g'> have the following con icts: r1 Scheme <predicate 1term 2>. Since <goal name g> is more preferable than <goal name g'>, I decided to commit to <goal name g>.
Rule maxUtility(g) ! pursued(g) r2 Scheme Since <goal name g> belonged to the set of goals that maximize the utility,
I decided to commit to <goal name g>.
Table 1: Examples of deliberation rules and explanations schemes Cycle : 004 (Why mop( p1 , p1 ) ) ? > mop( p1 , p1 ) and pickup ( p5 , p5 ) have the f o l l o w i n g c o n f l i c t s : r .
S i n c e mop( p1 , p1 ) i s more p r e f e r a b l e than pickup ( p5 , p5 ) , I d e c i d e d to commit to mop( p1 , p1 ) . > S i n c e mop( p1 , p1 ) b e l on g ed to the s e t o f g o a l s t h a t maximize the u t i l i t y , I d e c i d e d to commit to mop( p1 , p1 ) .
Fig. 1: Pseudo-natural language explanation for `Why mop(p1,p1)?'