AI Planning for Enterprise: Putting Theory Into Practice Shirin Sohrabi IBM Research ssohrab@us.ibm.com Abstract In this paper, I overview a number of AI Planning applications for Enterprise and discuss a number of challenges in applying AI Planning in that setting. I will also summarize the progress made to date in addressing these challenges. 1 Introduction Planning is a long-standing area of research within AI. Plan- ning is the task of finding a procedural course of action for a declaratively described system to reach its goals while opti- mizing overall performance measures. AI Planning can help when (1) your problem can be described in a declarative way; (2) you have domain knowledge that should not be ignored; (3) pure learning techniques are difficult to use either because there is a structure of the problem that cannot be learned by training or that there is little to no training data available; (4) you want to be able to explain a particular course of action the system took; or (5) you can leverage the existing relationships between a problem similar to yours to AI Planning. AI Planning has been applied to many applications includ- ing but not limited to robotics, manufacturing, logistics, trans- portation, and space 1 . I have in particular investigated the relationship between AI Planning and several applications, which I highlight below: • requirement engineering: helping analysts acquire a bet- ter understanding of the impact of high-level stakeholder preferences on low-level design decisions by modeling and reasoning with prioritization of stakeholder’s goal [Liaskos et al., 2010; Liaskos et al., 2011]; • web service composition: helping users customize the compositions of services with respect to their prefer- ences [Sohrabi and McIlraith, 2009; Sohrabi and McIl- raith, 2010; Sohrabi, 2012]; • stream processing; helping analyst with the task of au- tomating the composition of stream processing applica- tions [Sohrabi et al., 2013a]; • large-scale data analysis: helping users (network admin- istrators, nurses, physicians) by orchestrating the data 1 http://users.cecs.anu.edu.au/ ∼ patrik/sigaps/ analysis process automatically with a focus on two ap- plications: early detection of health complications in critical care, and detection of anomalous behaviors of network hosts in enterprise networks. In both cases, we generate multiple hypotheses in order to reason about possibly incomplete, noisy, or inconsistent sequences of observations received from external sources [Sohrabi et al., 2013b; Riabov et al., 2015]; • future state projection: providing analysts with the abil- ity to explain the given observations and based on that project multiple possible futures. Our focus was on en- ergy domain application where the objective is to project the price of oil and volume of oil produced 15 years into the future [Sohrabi et al., 2017]; and • enterprise risk management: assisting financial orga- nizations in identifying and managing emerging risks given observations derived from relevant news and so- cial media [Sohrabi et al., 2018b; Sohrabi et al., 2019]. A common theme among the applications that I was in- volved with is the connection to an established relation- ship to AI Planning. In particular, in my work I lever- age the existing relationship between diagnosis and plan- ning [Sohrabi et al., 2010], explanation generation as plan- ning [Sohrabi et al., 2011], and plan recognition as planning [Ram´ ırez and Geffner, 2009; Ram´ ırez and Geffner, 2010; Sohrabi et al., 2016a]. Furthermore, in these applications, the domain knowledge exists but not necessary in a form that is accessible by an AI planner; hence a major effort is required to address the knowledge engineering and extraction prob- lem. In addition, in applying AI Planning technologies we often encounter novel research problems. While, this novel problem is motivated by the application at hand, it has appli- cability and interest to the wider research community. In the next sections, I highlight these key research challenges and summarize the progress made to date in addressing them. 2 Relationship to Planning The first and fundamental challenge is defining the applica- tion problem and making the correspondence to AI Planning. That is to first define the computational problem, and then reduce it to a planning planning or to a problem in which an established relationship to AI Planning exists (e.g., plan Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19) 6408