One of the aims of DEVELOP is to apply artificial intelligence (AI) planning techniques to recommend learning opportunities for career development. More specifically, DEVELOP aims to identify potential sequences of learning interventions (e.g. eLearning courses, mentorship, changing teams) that will enable an individual to make progress towards the position(s) to which they aspire. The purpose of executing these learning interventions is to trigger desirable changes to the individual’s (transversal) competencies and/or social capital. The application of AI planning techniques will thus provide a personalised orchestration of existing learning interventions, offering a viable means for the individual to achieve their career goal(s). From the individual’s perspective, this will improve career awareness and facilitate internal mobility. For the organisation, this will optimise the use of existing learning resources, support talent management, and reduce attrition rates through improved employee engagement.
To some extent, previous research projects (e.g. WATCHME, EmployID, COLLAGE, PROLIX) and existing commercial products (e.g. Insala Career Development, SABA Talent Management) have attempted to address similar problems. For example, PROLIX recommends learning interventions based on the notion of a skills gap between an individual’s current competency profile and their desired competency profile, where learning interventions are treated independently (i.e. not as actions in a plan) and are assumed to have deterministic effects on competencies. Similarly, SABA’s TIM™ software offers a recommender system using machine learning techniques – similar to those employed by online retailers such as Amazon – but again treats each learning intervention independently and thus ignores the potentially complex interaction among learning interventions (e.g. dependencies, optimal sequences). To the best of our knowledge, no existing approach has attempted to address the problem holistically by recommending entire plans using AI planning techniques, with a formal treatment of uncertainty over the effects of learning interventions and optimality of the solution.
A guiding principle of DEVELOP’s approach further distinguishes it from existing work: that plan recommendations should be scrutable to the individual. That is, both the recommendations themselves and the reasons for the recommendations should be understandable to human users. This is essential for ensuring that an individual has trust in the system’s recommendations, and thus that the system is providing recommendations that are actually of benefit to the individual (i.e. to some extent, that the individual is likely to follow the recommendations). This principle has a number of implications on how we approach the problem. For example, while standard approaches to AI planning under uncertainty are capable of generating optimal contingency plans (i.e. optimal policies), such formulations are not easily understood by humans. Similarly, while there are many sophisticated AI planning techniques for generating optimal plans, few are capable of explaining why the plans are optimal.
Technically speaking, to solve these issues we first treat the problem as a non-observable Markov decision process (MDP) and apply conformant probabilistic planning techniques to generate good sequences of learning interventions to achieve a specified career goal. We will then apply probabilistic event reasoning and argumentation frameworks as a means of explaining the recommendations. The emphasis here is on generating simple plans that can be easily understood by humans in terms of possible trajectories to a goal, with the primary purpose being that of introspection. By introspection, we mean a process in which the individual can identify realistic career goals in terms of the probability that those goals can be achieved by various strategies, and the cost associated with those strategies. Similarly, it is hoped that this process will encourage the individual to reflect on their own preferences, to better understand how a career goal might be achieved, to identify sub-goals, and so on. Such information, along with historical information about the effects of learning interventions, can then be used to improve future plan recommendations via reinforcement learning.
Although simple sequences of learning interventions might fail (due to the potential for undesirable outcomes of learning interventions), they do offer an initial strategy for the individual to pursue; if and when the plan fails, online AI planning techniques can then be used to optimally pursue the original goal. Alternatively, new plan recommendations can be generated as part of a process of plan revision, e.g. the individual may choose to continue pursuing the original goal (making the process a form of replanning), or may choose a new goal to pursue. Thus, while we emphasise introspection, planning for the purpose of actually achieving a goal is still a fundamental concern.
Although DEVELOP aims to recommend plans for career development, it does not attempt to replace the expertise and insight of an experienced career guidance counsellor. It does, however, offer the potential to enhance career guidance counselling sessions with clear evidence of what plans might be better than others as a means of achieving a specific career goal.