A Deep Reinforcement Learning Framework with Formal Verification

Abstract

Artificial Intelligence (AI) and data are reshaping organizations and businesses. Human Resources (HR) management and talent development make no exception, as they tend to involve more automation and growing quantities of data. Because this brings implications on workforce, career transparency and equal opportunities, overseeing what fuels AI and analytical models, their quality standards, integrity and correctness becomes an imperative for those aspiring to such systems. Based on an ontology transformation to B-machines, this paper presents an approach to constructing a valid and error-free career agent with Deep Reinforcement Learning (DRL). In short, the agent's policy is built on a framework we called Multi State-Actor (MuStAc) using a decentralised training approach. Its purpose is to predict both relevant and valid career steps to employees, based on their profiles and company pathways (observations). Observations can comprise various data elements such as the current occupation, past experiences, performance, skills and qualifications, etc. The policy takes in all these observations and outputs the next recommended career step, in an environment set as the combination of a HR ontology and an Event-B model, which generates action spaces with respect to formal properties. The Event-B model and formal properties are derived using OWL to B transformation.

The Verification and Testing on Reactive Robot Cooperating Systems

“…In order to solve the problems often encountered in network structure design, such as lack of generalization ability and easy interference phenomenon, the growth process of the brain and the functional and structural characteristics of the biological nervous system are simulated first, a system and genetic algorithm are used to establish genetic search and artificial growth models, and then the network structure design is characterized by large search space and long training time. Inspired from previous studies [21,22,23,24,25,26,27] that combines machine learning with formal methods, we propose a new algorithm that first decomposes a complex problem into multiple simple problems, and then groups the subgroups corresponding to each subproblem. The processing units of the model are divided into three levels: primary processing unit, auxiliary processing unit and secondary processing unit.…”

The Verification and Testing on Reactive Robot Cooperating Systems

“…In order to solve the problems often encountered in network structure design, such as lack of generalization ability and easy interference phenomenon, the growth process of the brain and the functional and structural characteristics of the biological nervous system are simulated first, a system and genetic algorithm are used to establish genetic search and artificial growth models, and then the network structure design is characterized by large search space and long training time. Inspired from previous studies [21,22,23,24,25,26,27] that combines machine learning with formal methods, we propose a new algorithm that first decomposes a complex problem into multiple simple problems, and then groups the subgroups corresponding to each subproblem. The processing units of the model are divided into three levels: primary processing unit, auxiliary processing unit and secondary processing unit.…”