DSL Design for Reinforcement Learning Agents

Writing software that employs artificial intelligence (AI) is complex because the algorithms that must be implemented in general purpose programming languages are complex. One solution to this problem is to embed AI algorithms in domain specific languages (DSLs). DSLs are the “ultimate abstraction” for creating programs for a particular domain [1], but the question of how or even why to do this is not easily answered. We have developed a language with integrated reinforcement learning designed for writing intelligent agents. AFABL (A Friendly Adaptive Behavior Language), is implemented as an internal DSL shallowly embedded in the Scala programming language [3]. We discuss the development of AFABL, the basic elements of AFABL with an example, the way AFABL captures domain knowledge, the benefits of integrating reinforcement learning into a programming language and report the results of a programmer study which confirms and quantifies the usefulness of integrating reinforcement learning into a programming language. CCS Concepts • Computing methodologies→ Intelligent agents; Q-learning; • Software and its engineering → Domain specific languages; ACM Reference Format: Christopher Simpkins, Spencer Rugaber, and Charles Isbell, Jr.. 2017. DSL Design for Reinforcement Learning Agents. In Proceedings of Workshop on Domain-Specific Language Design and Implementation at SPLASH, Vancouver, Canada, October, 2017 (DSLDI-2017), 3 pages.