SySLLM: Generating Synthesized Policy Summaries for Reinforcement Learning Agents Using Large Language Models

Policies generated by Reinforcement Learning (RL) algorithms can be difficult to describe to users, as they result from the interplay between complex reward structures and neural network-based representations. This combination often leads to unpredictable behaviors, making policies challenging to analyze and posing significant obstacles to fostering human trust in real-world applications. Global policy summarization methods aim to describe agent behavior through a demonstration of actions in a subset of world-states. However, users can only watch a limited number of demonstrations, restricting their understanding of policies. Moreover, those methods overly rely on user interpretation, as they do not synthesize observations into coherent patterns. In this work, we present SySLLM (Synthesized Summary using LLMs), a novel method that employs synthesis summarization, utilizing large language models' (LLMs) extensive world knowledge and ability to capture patterns, to generate textual summaries of policies. Specifically, an expert evaluation demonstrates that the proposed approach generates summaries that capture the main insights generated by experts while not resulting in significant hallucinations. Additionally, a user study shows that SySLLM summaries are preferred over demonstration-based policy summaries and match or surpass their performance in objective agent identification tasks.

Envious Explore and Exploit

Explore-and-exploit tradeoffs play a key role in recommendation systems (RSs), aiming at serving users better by learning from previous interactions. Despite their commercial success, the societal effects of explore-and-exploit mechanisms are not well understood, especially regarding the utility discrepancy they generate between different users. In this work, we measure such discrepancy using the economic notion of envy. We present a multi-armed bandit-like model in which every round consists of several sessions, and rewards are realized once per round. We call the latter property reward consistency, and show that the RS can leverage this property for better societal outcomes. On the downside, doing so also generates envy, as late-to-arrive users enjoy the information gathered by early-to-arrive users. We examine the generated envy under several arrival order mechanisms and virtually any anonymous algorithm, i.e., any algorithm that treats all similar users similarly without leveraging their identities. We provide tight envy bounds on uniform arrival and upper bound the envy for nudged arrival, in which the RS can affect the order of arrival by nudging its users. Furthermore, we study the efficiency-fairness trade-off by devising an algorithm that allows constant envy and approximates the optimal welfare in restricted settings. Finally, we validate our theoretical results empirically using simulations.

Selective Response Strategies for GenAI

The rise of Generative AI (GenAI) has significantly impacted human-based forums like Stack Overflow, which are essential for generating high-quality data. This creates a negative feedback loop, hindering the development of GenAI systems, which rely on such data to provide accurate responses. In this paper, we provide a possible remedy: A novel strategy we call selective response. Selective response implies that GenAI could strategically provide inaccurate (or conservative) responses to queries involving emerging topics and novel technologies, thereby driving users to use human-based forums like Stack Overflow. We show that selective response can potentially have a compounding effect on the data generation process, increasing both GenAI's revenue and user welfare in the long term. From an algorithmic perspective, we propose an approximately optimal approach to maximize GenAI's revenue under social welfare constraints. From a regulatory perspective, we derive sufficient and necessary conditions for selective response to improve welfare improvements.

Personalized Reinforcement Learning with a Budget of Policies

Personalization in machine learning (ML) tailors models' decisions to the individual characteristics of users. While this approach has seen success in areas like recommender systems, its expansion into high-stakes fields such as healthcare and autonomous driving is hindered by the extensive regulatory approval processes involved. To address this challenge, we propose a novel framework termed represented Markov Decision Processes (r-MDPs) that is designed to balance the need for personalization with the regulatory constraints. In an r-MDP, we cater to a diverse user population, each with unique preferences, through interaction with a small set of representative policies. Our objective is twofold: efficiently match each user to an appropriate representative policy and simultaneously optimize these policies to maximize overall social welfare. We develop two deep reinforcement learning algorithms that efficiently solve r-MDPs. These algorithms draw inspiration from the principles of classic K-means clustering and are underpinned by robust theoretical foundations. Our empirical investigations, conducted across a variety of simulated environments, showcase the algorithms' ability to facilitate meaningful personalization even under constrained policy budgets. Furthermore, they demonstrate scalability, efficiently adapting to larger policy budgets.

Principal-Agent Reward Shaping in MDPs

Principal-agent problems arise when one party acts on behalf of another, leading to conflicts of interest. The economic literature has extensively studied principal-agent problems, and recent work has extended this to more complex scenarios such as Markov Decision Processes (MDPs). In this paper, we further explore this line of research by investigating how reward shaping under budget constraints can improve the principal's utility. We study a two-player Stackelberg game where the principal and the agent have different reward functions, and the agent chooses an MDP policy for both players. The principal offers an additional reward to the agent, and the agent picks their policy selfishly to maximize their reward, which is the sum of the original and the offered reward. Our results establish the NP-hardness of the problem and offer polynomial approximation algorithms for two classes of instances: Stochastic trees and deterministic decision processes with a finite horizon.

Learning with Exposure Constraints in Recommendation Systems

Recommendation systems are dynamic economic systems that balance the needs of multiple stakeholders. A recent line of work studies incentives from the content providers' point of view. Content providers, e.g., vloggers and bloggers, contribute fresh content and rely on user engagement to create revenue and finance their operations. In this work, we propose a contextual multi-armed bandit setting to model the dependency of content providers on exposure. In our model, the system receives a user context in every round and has to select one of the arms. Every arm is a content provider who must receive a minimum number of pulls every fixed time period (e.g., a month) to remain viable in later rounds; otherwise, the arm departs and is no longer available. The system aims to maximize the users' (content consumers) welfare. To that end, it should learn which arms are vital and ensure they remain viable by subsidizing arm pulls if needed. We develop algorithms with sub-linear regret, as well as a lower bound that demonstrates that our algorithms are optimal up to logarithmic factors.

Modeling Attrition in Recommender Systems with Departing Bandits

Traditionally, when recommender systems are formalized as multi-armed bandits, the policy of the recommender system influences the rewards accrued, but not the length of interaction. However, in real-world systems, dissatisfied users may depart (and never come back). In this work, we propose a novel multi-armed bandit setup that captures such policy-dependent horizons. Our setup consists of a finite set of user types, and multiple arms with Bernoulli payoffs. Each (user type, arm) tuple corresponds to an (unknown) reward probability. Each user's type is initially unknown and can only be inferred through their response to recommendations. Moreover, if a user is dissatisfied with their recommendation, they might depart the system. We first address the case where all users share the same type, demonstrating that a recent UCB-based algorithm is optimal. We then move forward to the more challenging case, where users are divided among two types. While naive approaches cannot handle this setting, we provide an efficient learning algorithm that achieves $\tilde{O}(\sqrt{T})$ regret, where $T$ is the number of users.

Optimizing Long-term Social Welfare in Recommender Systems: A Constrained Matching Approach

Most recommender systems (RS) research assumes that a user's utility can be maximized independently of the utility of the other agents (e.g., other users, content providers). In realistic settings, this is often not true---the dynamics of an RS ecosystem couple the long-term utility of all agents. In this work, we explore settings in which content providers cannot remain viable unless they receive a certain level of user engagement. We formulate the recommendation problem in this setting as one of equilibrium selection in the induced dynamical system, and show that it can be solved as an optimal constrained matching problem. Our model ensures the system reaches an equilibrium with maximal social welfare supported by a sufficiently diverse set of viable providers. We demonstrate that even in a simple, stylized dynamical RS model, the standard myopic approach to recommendation---always matching a user to the best provider---performs poorly. We develop several scalable techniques to solve the matching problem, and also draw connections to various notions of user regret and fairness, arguing that these outcomes are fairer in a utilitarian sense.

Learning under Invariable Bayesian Safety

A recent body of work addresses safety constraints in explore-and-exploit systems. Such constraints arise where, for example, exploration is carried out by individuals whose welfare should be balanced with overall welfare. In this paper, we adopt a model inspired by recent work on a bandit-like setting for recommendations. We contribute to this line of literature by introducing a safety constraint that should be respected in every round and determines that the expected value in each round is above a given threshold. Due to our modeling, the safe explore-and-exploit policy deserves careful planning, or otherwise, it will lead to sub-optimal welfare. We devise an asymptotically optimal algorithm for the setting and analyze its instance-dependent convergence rate.

Predicting Strategic Behavior from Free Text

The connection between messaging and action is fundamental both to web applications, such as web search and sentiment analysis, and to economics. However, while prominent online applications exploit messaging in natural (human) language in order to predict non-strategic action selection, the economics literature focuses on the connection between structured stylized messaging to strategic decisions in games and multi-agent encounters. This paper aims to connect these two strands of research, which we consider highly timely and important due to the vast online textual communication on the web. Particularly, we introduce the following question: can free text expressed in natural language serve for the prediction of action selection in an economic context, modeled as a game? In order to initiate the research on this question, we introduce the study of an individual's action prediction in a one-shot game based on free text he/she provides, while being unaware of the game to be played. We approach the problem by attributing commonsensical personality attributes via crowd-sourcing to free texts written by individuals, and employing transductive learning to predict actions taken by these individuals in one-shot games based on these attributes. Our approach allows us to train a single classifier that can make predictions with respect to actions taken in multiple games. In experiments with three well-studied games, our algorithm compares favorably with strong alternative approaches. In ablation analysis, we demonstrate the importance of our modeling choices -- the representation of the text with the commonsensical personality attributes and our classifier -- to the predictive power of our model.