Pearl: A Production-ready Reinforcement Learning Agent

Reinforcement Learning (RL) offers a versatile framework for achieving long-term goals. Its generality allows us to formalize a wide range of problems that real-world intelligent systems encounter, such as dealing with delayed rewards, handling partial observability, addressing the exploration and exploitation dilemma, utilizing offline data to improve online performance, and ensuring safety constraints are met. Despite considerable progress made by the RL research community in addressing these issues, existing open-source RL libraries tend to focus on a narrow portion of the RL solution pipeline, leaving other aspects largely unattended. This paper introduces Pearl, a Production-ready RL agent software package explicitly designed to embrace these challenges in a modular fashion. In addition to presenting preliminary benchmark results, this paper highlights Pearl's industry adoptions to demonstrate its readiness for production usage. Pearl is open sourced on Github at github.com/facebookresearch/pearl and its official website is located at pearlagent.github.io.

Optimizing Long-term Value for Auction-Based Recommender Systems via On-Policy Reinforcement Learning

Auction-based recommender systems are prevalent in online advertising platforms, but they are typically optimized to allocate recommendation slots based on immediate expected return metrics, neglecting the downstream effects of recommendations on user behavior. In this study, we employ reinforcement learning to optimize for long-term return metrics in an auction-based recommender system. Utilizing temporal difference learning, a fundamental reinforcement learning algorithm, we implement an one-step policy improvement approach that biases the system towards recommendations with higher long-term user engagement metrics. This optimizes value over long horizons while maintaining compatibility with the auction framework. Our approach is grounded in dynamic programming ideas which show that our method provably improves upon the existing auction-based base policy. Through an online A/B test conducted on an auction-based recommender system which handles billions of impressions and users daily, we empirically establish that our proposed method outperforms the current production system in terms of long-term user engagement metrics.

AMOM: Adaptive Masking over Masking for Conditional Masked Language Model

Transformer-based autoregressive (AR) methods have achieved appealing performance for varied sequence-to-sequence generation tasks, e.g., neural machine translation, summarization, and code generation, but suffer from low inference efficiency. To speed up the inference stage, many non-autoregressive (NAR) strategies have been proposed in the past few years. Among them, the conditional masked language model (CMLM) is one of the most versatile frameworks, as it can support many different sequence generation scenarios and achieve very competitive performance on these tasks. In this paper, we further introduce a simple yet effective adaptive masking over masking strategy to enhance the refinement capability of the decoder and make the encoder optimization easier. Experiments on \textbf{3} different tasks (neural machine translation, summarization, and code generation) with \textbf{15} datasets in total confirm that our proposed simple method achieves significant performance improvement over the strong CMLM model. Surprisingly, our proposed model yields state-of-the-art performance on neural machine translation (\textbf{34.62} BLEU on WMT16 EN$\to$RO, \textbf{34.82} BLEU on WMT16 RO$\to$EN, and \textbf{34.84} BLEU on IWSLT De$\to$En) and even better performance than the \textbf{AR} Transformer on \textbf{7} benchmark datasets with at least \textbf{2.2$\times$} speedup. Our code is available at GitHub.

A Validation Tool for Designing Reinforcement Learning Environments

Reinforcement learning (RL) has gained increasing attraction in the academia and tech industry with launches to a variety of impactful applications and products. Although research is being actively conducted on many fronts (e.g., offline RL, performance, etc.), many RL practitioners face a challenge that has been largely ignored: determine whether a designed Markov Decision Process (MDP) is valid and meaningful. This study proposes a heuristic-based feature analysis method to validate whether an MDP is well formulated. We believe an MDP suitable for applying RL should contain a set of state features that are both sensitive to actions and predictive in rewards. We tested our method in constructed environments showing that our approach can identify certain invalid environment formulations. As far as we know, performing validity analysis for RL problem formulation is a novel direction. We envision that our tool will serve as a motivational example to help practitioners apply RL in real-world problems more easily.

Semantic Parsing Natural Language into Relational Algebra

Natural interface to database (NLIDB) has been researched a lot during the past decades. In the core of NLIDB, is a semantic parser used to convert natural language into SQL. Solutions from traditional NLP methodology focuses on grammar rule pattern learning and pairing via intermediate logic forms. Although those methods give an acceptable performance on certain specific database and parsing tasks, they are hard to generalize and scale. On the other hand, recent progress in neural deep learning seems to provide a promising direction towards building a general NLIDB system. Unlike the traditional approach, those neural methodologies treat the parsing problem as a sequence-to-sequence learning problem. In this paper, we experimented on several sequence-to-sequence learning models and evaluate their performance on general database parsing task.

Dual Monte Carlo Tree Search

AlphaZero, using a combination of Deep Neural Networks and Monte Carlo Tree Search (MCTS), has successfully trained reinforcement learning agents in a tabula-rasa way. The neural MCTS algorithm has been successful in finding near-optimal strategies for games through self-play. However, the AlphaZero algorithm has a significant drawback; it takes a long time to converge and requires high computational power due to complex neural networks for solving games like Chess, Go, Shogi, etc. Owing to this, it is very difficult to pursue neural MCTS research without cutting-edge hardware, which is a roadblock for many aspiring neural MCTS researchers. In this paper, we propose a new neural MCTS algorithm, called Dual MCTS, which helps overcome these drawbacks. Dual MCTS uses two different search trees, a single deep neural network, and a new update technique for the search trees using a combination of the PUCB, a sliding-window, and the epsilon-greedy algorithm. This technique is applicable to any MCTS based algorithm to reduce the number of updates to the tree. We show that Dual MCTS performs better than one of the most widely used neural MCTS algorithms, AlphaZero, for various symmetric and asymmetric games.

Solving QSAT problems with neural MCTS

Recent achievements from AlphaZero using self-play has shown remarkable performance on several board games. It is plausible to think that self-play, starting from zero knowledge, can gradually approximate a winning strategy for certain two-player games after an amount of training. In this paper, we try to leverage the computational power of neural Monte Carlo Tree Search (neural MCTS), the core algorithm from AlphaZero, to solve Quantified Boolean Formula Satisfaction (QSAT) problems, which are PSPACE complete. Knowing that every QSAT problem is equivalent to a QSAT game, the game outcome can be used to derive the solutions of the original QSAT problems. We propose a way to encode Quantified Boolean Formulas (QBFs) as graphs and apply a graph neural network (GNN) to embed the QBFs into the neural MCTS. After training, an off-the-shelf QSAT solver is used to evaluate the performance of the algorithm. Our result shows that, for problems within a limited size, the algorithm learns to solve the problem correctly merely from self-play.

First-Order Problem Solving through Neural MCTS based Reinforcement Learning

The formal semantics of an interpreted first-order logic (FOL) statement can be given in Tarskian Semantics or a basically equivalent Game Semantics. The latter maps the statement and the interpretation into a two-player semantic game. Many combinatorial problems can be described using interpreted FOL statements and can be mapped into a semantic game. Therefore, learning to play a semantic game perfectly leads to the solution of a specific instance of a combinatorial problem. We adapt the AlphaZero algorithm so that it becomes better at learning to play semantic games that have different characteristics than Go and Chess. We propose a general framework, Persephone, to map the FOL description of a combinatorial problem to a semantic game so that it can be solved through a neural MCTS based reinforcement learning algorithm. Our goal for Persephone is to make it tabula-rasa, mapping a problem stated in interpreted FOL to a solution without human intervention.

Learning Self-Game-Play Agents for Combinatorial Optimization Problems

Recent progress in reinforcement learning (RL) using self-game-play has shown remarkable performance on several board games (e.g., Chess and Go) as well as video games (e.g., Atari games and Dota2). It is plausible to consider that RL, starting from zero knowledge, might be able to gradually approximate a winning strategy after a certain amount of training. In this paper, we explore neural Monte-Carlo-Tree-Search (neural MCTS), an RL algorithm which has been applied successfully by DeepMind to play Go and Chess at a super-human level. We try to leverage the computational power of neural MCTS to solve a class of combinatorial optimization problems. Following the idea of Hintikka's Game-Theoretical Semantics, we propose the Zermelo Gamification (ZG) to transform specific combinatorial optimization problems into Zermelo games whose winning strategies correspond to the solutions of the original optimization problem. The ZG also provides a specially designed neural MCTS. We use a combinatorial planning problem for which the ground-truth policy is efficiently computable to demonstrate that ZG is promising.