Search

This paper develops theory and algorithms for interacting large language model agents (LLMAs) using methods from statistical signal processing and microeconomics. While both fields are mature, their application to decision-making by interacting LLMAs remains unexplored. Motivated by Bayesian sentiment analysis on online platforms, we construct interpretable models and stochastic control algorithms that enable LLMAs to interact and perform Bayesian inference. Because interacting LLMAs learn from prior decisions and external inputs, they exhibit bias and herding behavior. Thus, developing interpretable models and stochastic control algorithms is essential to understand and mitigate these behaviors. This paper has three main results. First, we show using Bayesian revealed preferences from microeconomics that an individual LLMA satisfies the sufficient conditions for rationally inattentive (bounded rationality) utility maximization and, given an observation, the LLMA chooses an action that maximizes a regularized utility. Second, we utilize Bayesian social learning to construct interpretable models for LLMAs that interact sequentially with each other and the environment while performing Bayesian inference. Our models capture the herding behavior exhibited by interacting LLMAs. Third, we propose a stochastic control framework to delay herding and improve state estimation accuracy under two settings: (a) centrally controlled LLMAs and (b) autonomous LLMAs with incentives. Throughout the paper, we demonstrate the efficacy of our methods on real datasets for hate speech classification and product quality assessment, using open-source models like Mistral and closed-source models like ChatGPT. The main takeaway of this paper, based on substantial empirical analysis and mathematical formalism, is that LLMAs act as rationally bounded Bayesian agents that exhibit social learning when interacting.

This paper considers the problem of annotating datapoints using an expert with only a few annotation rounds in a label-scarce setting. We propose soliciting reliable feedback on difficulty in annotating a datapoint from the expert in addition to ground truth label. Existing literature in active learning or coreset selection turns out to be less relevant to our setting since they presume the existence of a reliable trained model, which is absent in the label-scarce regime. However, the literature on coreset selection emphasizes the presence of difficult data points in the training set to perform supervised learning in downstream tasks (Mindermann et al., 2022). Therefore, for a given fixed annotation budget of $\mathsf{T}$ rounds, we model the sequential decision-making problem of which (difficult) datapoints to choose for annotation in a sparse linear bandits framework with the constraint that no arm can be pulled more than once (blocking constraint). With mild assumptions on the datapoints, our (computationally efficient) Explore-Then-Commit algorithm BSLB achieves a regret guarantee of $\widetilde{\mathsf{O}}(k^{\frac{1}{3}} \mathsf{T}^{\frac{2}{3}} +k^{-\frac{1}{2}} \beta_k + k^{-\frac{1}{12}} \beta_k^{\frac{1}{2}}\mathsf{T}^{\frac{5}{6}})$ where the unknown parameter vector has tail magnitude $\beta_k$ at sparsity level $k$. To this end, we show offline statistical guarantees of Lasso estimator with mild Restricted Eigenvalue (RE) condition that is also robust to sparsity. Finally, we propose a meta-algorithm C-BSLB that does not need knowledge of the optimal sparsity parameters at a no-regret cost. We demonstrate the efficacy of our BSLB algorithm for annotation in the label-scarce setting for an image classification task on the PASCAL-VOC dataset, where we use real-world annotation difficulty scores., Comment: 31 Pages

We analyze the finite sample regret of a decreasing step size stochastic gradient algorithm. We assume correlated noise and use a perturbed Lyapunov function as a systematic approach for the analysis. Finally we analyze the escape time of the iterates using large deviations theory.

We consider word-of-mouth social learning involving $m$ Kalman filter agents that operate sequentially. The first Kalman filter receives the raw observations, while each subsequent Kalman filter receives a noisy measurement of the conditional mean of the previous Kalman filter. The prior is updated by the $m$-th Kalman filter. When $m=2$, and the observations are noisy measurements of a Gaussian random variable, the covariance goes to zero as $k^{-1/3}$ for $k$ observations, instead of $O(k^{-1})$ in the standard Kalman filter. In this paper we prove that for $m$ agents, the covariance decreases to zero as $k^{-(2^m-1)}$, i.e, the learning slows down exponentially with the number of agents. We also show that by artificially weighing the prior at each time, the learning rate can be made optimal as $k^{-1}$. The implication is that in word-of-mouth social learning, artificially re-weighing the prior can yield the optimal learning rate.

We derive a minimax distributionally robust inverse reinforcement learning (IRL) algorithm to reconstruct the utility functions of a multi-agent sensing system. Specifically, we construct utility estimators which minimize the worst-case prediction error over a Wasserstein ambiguity set centered at noisy signal observations. We prove the equivalence between this robust estimation and a semi-infinite optimization reformulation, and we propose a consistent algorithm to compute solutions. We illustrate the efficacy of this robust IRL scheme in numerical studies to reconstruct the utility functions of a cognitive radar network from observed tracking signals.

Motivated by learning of correlated equilibria in non-cooperative games, we perform a large deviations analysis of a regret minimizing stochastic approximation algorithm. The regret minimization algorithm we consider comprises multiple agents that communicate over a graph to coordinate their decisions. We derive an exponential decay rate towards the algorithm's stable point using large deviations theory. Our analysis leverages the variational representation of the Laplace functionals and weak convergence methods to characterize the exponential decay rate.

This paper studies the problem of autonomous agents performing Bayesian social learning for sequential detection when the observations of the state belong to a high-dimensional space and are expensive to analyze. Specifically, when the observations are textual, the Bayesian agent can use a large language model (LLM) as a map to get a low-dimensional private observation. The agent performs Bayesian learning and takes an action that minimizes the expected cost and is visible to subsequent agents. We prove that a sequence of such Bayesian agents herd in finite time to the public belief and take the same action disregarding the private observations. We propose a stopping time formulation for quickest time herding in social learning and optimally balance privacy and herding. Structural results are shown on the threshold nature of the optimal policy to the stopping time problem. We illustrate the application of our framework when autonomous Bayesian detectors aim to sequentially identify if a user is a hate speech peddler on an online platform by parsing text observations using an LLM. We numerically validate our results on real-world hate speech datasets. We show that autonomous Bayesian agents designed to flag hate speech peddlers in online platforms herd and misclassify the users when the public prior is strong. We also numerically show the effect of a threshold policy in delaying herding.

This paper studies how a stochastic gradient algorithm (SG) can be controlled to hide the estimate of the local stationary point from an eavesdropper. Such problems are of significant interest in distributed optimization settings like federated learning and inventory management. A learner queries a stochastic oracle and incentivizes the oracle to obtain noisy gradient measurements and perform SG. The oracle probabilistically returns either a noisy gradient of the function} or a non-informative measurement, depending on the oracle state and incentive. The learner's query and incentive are visible to an eavesdropper who wishes to estimate the stationary point. This paper formulates the problem of the learner performing covert optimization by dynamically incentivizing the stochastic oracle and obfuscating the eavesdropper as a finite-horizon Markov decision process (MDP). Using conditions for interval-dominance on the cost and transition probability structure, we show that the optimal policy for the MDP has a monotone threshold structure. We propose searching for the optimal stationary policy with the threshold structure using a stochastic approximation algorithm and a multi-armed bandit approach. The effectiveness of our methods is numerically demonstrated on a covert federated learning hate-speech classification task.

This paper constructs an algorithmic framework for adaptively achieving the mechanism design objective, finding a mechanism inducing socially optimal Nash equilibria, without knowledge of the utility functions of the agents. We consider a probing scheme where the designer can iteratively enact mechanisms and observe Nash equilibria responses. We first derive necessary and sufficient conditions, taking the form of linear program feasibility, for the existence of utility functions under which the empirical Nash equilibria responses are socially optimal. Then, we utilize this to construct a loss function with respect to the mechanism, and show that its global minimization occurs at mechanisms under which Nash equilibria system responses are also socially optimal. We develop a simulated annealing-based gradient algorithm, and prove that it converges in probability to this set of global minima, thus achieving adaptive mechanism design.

This paper introduces a novel non-parametric change detection algorithm to identify eclipse attacks on a blockchain network; the non-parametric algorithm relies only on the empirical mean and variance of the dataset, making it highly adaptable. An eclipse attack occurs when malicious actors isolate blockchain users, disrupting their ability to reach consensus with the broader network, thereby distorting their local copy of the ledger. To detect an eclipse attack, we monitor changes in the Fr\'echet mean and variance of the evolving blockchain communication network connecting blockchain users. First, we leverage the Johnson-Lindenstrauss lemma to project large-dimensional networks into a lower-dimensional space, preserving essential statistical properties. Subsequently, we employ a non-parametric change detection procedure, leading to a test statistic that converges weakly to a Brownian bridge process in the absence of an eclipse attack. This enables us to quantify the false alarm rate of the detector. Our detector can be implemented as a smart contract on the blockchain, offering a tamper-proof and reliable solution. Finally, we use numerical examples to compare the proposed eclipse attack detector with a detector based on the random forest model.

How to design a Markov Decision Process (MDP) based radar controller that makes small sacrifices in performance to mask its sensing plan from an adversary? The radar controller purposefully minimizes the Fisher information of its emissions so that an adversary cannot identify the controller's model parameters accurately. Unlike classical open loop statistical inference, where the Fisher information serves as a lower bound for the achievable covariance, this paper employs the Fisher information as a design constraint for a closed loop radar controller to mask its sensing plan. We analytically derive a closed-form expression for the determinant of the Fisher Information Matrix (FIM) pertaining to the parameters of the MDP-based controller. Subsequently, we constrain the MDP with respect to the determinant of the FIM. Numerical results show that the introduction of minor perturbations to the MDP's transition kernel and the total operation cost can reduce the Fisher Information of the emissions. Consequently, this reduction amplifies the variability in policy and transition kernel estimation errors, thwarting the adversary's accuracy in estimating the controller's sensing plan.

A learner aims to minimize a function $f$ by repeatedly querying a distributed oracle that provides noisy gradient evaluations. At the same time, the learner seeks to hide $\arg\min f$ from a malicious eavesdropper that observes the learner's queries. This paper considers the problem of \textit{covert} or \textit{learner-private} optimization, where the learner has to dynamically choose between learning and obfuscation by exploiting the stochasticity. The problem of controlling the stochastic gradient algorithm for covert optimization is modeled as a Markov decision process, and we show that the dynamic programming operator has a supermodular structure implying that the optimal policy has a monotone threshold structure. A computationally efficient policy gradient algorithm is proposed to search for the optimal querying policy without knowledge of the transition probabilities. As a practical application, our methods are demonstrated on a hate speech classification task in a federated setting where an eavesdropper can use the optimal weights to generate toxic content, which is more easily misclassified. Numerical results show that when the learner uses the optimal policy, an eavesdropper can only achieve a validation accuracy of $52\%$ with no information and $69\%$ when it has a public dataset with 10\% positive samples compared to $83\%$ when the learner employs a greedy policy.

This paper proposes a new approach for change point detection in causal networks of multivariate Hawkes processes using Frechet statistics. Our method splits the point process into overlapping windows, estimates kernel matrices in each window, and reconstructs the signed Laplacians by treating the kernel matrices as the adjacency matrices of the causal network. We demonstrate the effectiveness of our method through experiments on both simulated and real-world cryptocurrency datasets. Our results show that our method is capable of accurately detecting and characterizing changes in the causal structure of multivariate Hawkes processes, and may have potential applications in fields such as finance and neuroscience. The proposed method is an extension of previous work on Frechet statistics in point process settings and represents an important contribution to the field of change point detection in multivariate point processes.

What are the optimal times for an Internet of Things (IoT) device to act as a blockchain miner? The aim is to minimize the energy consumed by low-power IoT devices that log their data into a secure (tamper-proof) distributed ledger. We formulate a multiple stopping time Bayesian sequential detection problem to address energy-efficient blockchain mining for IoT devices. The objective is to identify $L$ optimal stops for mining, thereby maximizing the probability of successfully adding a block to the blockchain; we also present a model to optimize the number of stops (mining instants). The formulation is equivalent to a multiple stopping time POMDP. Since POMDPs are in general computationally intractable to solve, we show mathematically using submodularity arguments that the optimal mining policy has a useful structure: 1) it is monotone in belief space, and 2) it exhibits a threshold structure, which divides the belief space into two connected sets. Exploiting the structural results, we formulate a computationally-efficient linear mining policy for the blockchain-enabled IoT device. We present a policy gradient technique to optimize the parameters of the linear mining policy. Finally, we use synthetic and real Bitcoin datasets to study the performance of our proposed mining policy. We demonstrate the energy efficiency achieved by the optimal linear mining policy in contrast to other heuristic strategies.

Consider a target being tracked by a cognitive radar network. If the target can intercept noisy radar emissions, how can it detect coordination in the radar network? By 'coordination' we mean that the radar emissions satisfy Pareto optimality with respect to multi-objective optimization over the objective functions of each radar and a constraint on total network power output. This paper provides a novel inverse multi-objective optimization approach for statistically detecting Pareto optimal ('coordinating') behavior, from a finite dataset of noisy radar emissions. Specifically, we develop necessary and sufficient conditions for radar network emissions to be consistent with multi-objective optimization (coordination), and we provide a statistical detector with theoretical guarantees for determining this consistency when radar emissions are observed in noise. We also provide numerical simulations which validate our approach. Note that while we make use of the specific framework of a radar network coordination problem, our results apply more generally to the field of inverse multi-objective optimization.

This paper provides a finite-sample analysis of a passive stochastic gradient Langevin dynamics algorithm (PSGLD) designed to achieve adaptive inverse reinforcement learning (IRL). By passive, we mean that the noisy gradients available to the PSGLD algorithm (inverse learning process) are evaluated at randomly chosen points by an external stochastic gradient algorithm (forward learner) that aims to optimize a cost function. The PSGLD algorithm acts as a randomized sampler to achieve adaptive IRL by reconstructing this cost function nonparametrically from the stationary measure of a Langevin diffusion. Previous work has analyzed the asymptotic performance of this passive algorithm using weak convergence techniques. This paper analyzes the non-asymptotic (finite-sample) performance using a logarithmic-Sobolev inequality and the Otto-Villani Theorem. We obtain finite-sample bounds on the 2-Wasserstein distance between the estimates generated by the PSGLD algorithm and the cost function. Apart from achieving finite-sample guarantees for adaptive IRL, this work extends a line of research in analysis of passive stochastic gradient algorithms to the finite-sample regime for Langevin dynamics.

This study presents a novel deep learning method, called GATv2-GCN, for predicting player performance in sports. To construct a dynamic player interaction graph, we leverage player statistics and their interactions during gameplay. We use a graph attention network to capture the attention that each player pays to each other, allowing for more accurate modeling of the dynamic player interactions. To handle the multivariate player statistics time series, we incorporate a temporal convolution layer, which provides the model with temporal predictive power. We evaluate the performance of our model using real-world sports data, demonstrating its effectiveness in predicting player performance. Furthermore, we explore the potential use of our model in a sports betting context, providing insights into profitable strategies that leverage our predictive power. The proposed method has the potential to advance the state-of-the-art in player performance prediction and to provide valuable insights for sports analytics and betting industries.

This paper proposes a method to detect change points in dynamic social networks using Fr\'echet statistics. We address two main questions: (1) what metric can quantify the distances between graph Laplacians in a dynamic network and enable efficient computation, and (2) how can the Fr\'echet statistics be extended to detect multiple change points while maintaining the significance level of the hypothesis test? Our solution defines a metric space for graph Laplacians using the Log-Euclidean metric, enabling a closed-form formula for Fr\'echet mean and variance. We present a framework for change point detection using Fr\'echet statistics and extend it to multiple change points with binary segmentation. The proposed algorithm uses incremental computation for Fr\'echet mean and variance to improve efficiency and is validated on simulated and two real-world datasets, namely the UCI message dataset and the Enron email dataset.

We propose a new way to measure inequalities such as the glass ceiling effect in attributed networks. Existing measures typically rely solely on node degree distribution or degree assortativity, but our approach goes beyond these measures by using mutual information (based on Shannon and more generally, Renyi entropy) between the conditional probability distributions of node attributes given node degrees of adjacent nodes. We show that this mutual information measure aligns with both the analytical structural inequality model and historical publication data, making it a reliable approach to capture the complexities of attributed networks. Specifically, we demonstrate this through an analysis of citation networks. Moreover, we propose a stochastic optimization algorithm using a parameterized conditional logit model for edge addition, which outperforms a baseline uniform distribution. By recommending links at random using this algorithm, we can mitigate the glass ceiling effect, which is a crucial tool in addressing structural inequalities in networks.

In this paper, we exploit the spiked covariance structure of the clutter plus noise covariance matrix for radar signal processing. Using state-of-the-art techniques high dimensional statistics, we propose a nonlinear shrinkage-based rotation invariant spiked covariance matrix estimator. We state the convergence of the estimated spiked eigenvalues. We use a dataset generated from the high-fidelity, site-specific physics-based radar simulation software RFView to compare the proposed algorithm against the existing Rank Constrained Maximum Likelihood (RCML)-Expected Likelihood (EL) covariance estimation algorithm. We demonstrate that the computation time for the estimation by the proposed algorithm is less than the RCML-EL algorithm with identical Signal to Clutter plus Noise (SCNR) performance. We show that the proposed algorithm and the RCML-EL-based algorithm share the same optimization problem in high dimensions. We use Low-Rank Adaptive Normalized Matched Filter (LR-ANMF) detector to compute the detection probabilities for different false alarm probabilities over a range of target SNR. We present preliminary results which demonstrate the robustness of the detector against contaminating clutter discretes using the Challenge Dataset from RFView. Finally, we empirically show that the minimum variance distortionless beamformer (MVDR) error variance for the proposed algorithm is identical to the error variance resulting from the true covariance matrix.

This paper surveys mathematical models, structural results and algorithms in controlled sensing with social learning in social networks. Part 1, namely Bayesian Social Learning with Controlled Sensing addresses the following questions: How does risk averse behavior in social learning affect quickest change detection? How can information fusion be priced? How is the convergence rate of state estimation affected by social learning? The aim is to develop and extend structural results in stochastic control and Bayesian estimation to answer these questions. Such structural results yield fundamental bounds on the optimal performance, give insight into what parameters affect the optimal policies, and yield computationally efficient algorithms. Part 2, namely, Multi-agent Information Fusion with Behavioral Economics Constraints generalizes Part 1. The agents exhibit sophisticated decision making in a behavioral economics sense; namely the agents make anticipatory decisions (thus the decision strategies are time inconsistent and interpreted as subgame Bayesian Nash equilibria). Part 3, namely {\em Interactive Sensing in Large Networks}, addresses the following questions: How to track the degree distribution of an infinite random graph with dynamics (via a stochastic approximation on a Hilbert space)? How can the infected degree distribution of a Markov modulated power law network and its mean field dynamics be tracked via Bayesian filtering given incomplete information obtained by sampling the network? We also briefly discuss how the glass ceiling effect emerges in social networks. Part 4, namely \emph{Efficient Network Polling} deals with polling in large scale social networks. In such networks, only a fraction of nodes can be polled to determine their decisions. Which nodes should be polled to achieve a statistically accurate estimates?

This paper considers adaptive radar electronic counter-counter measures (ECCM) to mitigate ECM by an adversarial jammer. Our ECCM approach models the jammer-radar interaction as a Principal Agent Problem (PAP), a popular economics framework for interaction between two entities with an information imbalance. In our setup, the radar does not know the jammer's utility. Instead, the radar learns the jammer's utility adaptively over time using inverse reinforcement learning. The radar's adaptive ECCM objective is two-fold (1) maximize its utility by solving the PAP, and (2) estimate the jammer's utility by observing its response. Our adaptive ECCM scheme uses deep ideas from revealed preference in micro-economics and principal agent problem in contract theory. Our numerical results show that, over time, our adaptive ECCM both identifies and mitigates the jammer's utility.

Consider a target being tracked by a cognitive radar network. If the target can intercept some radar network emissions, how can it detect coordination among the radars? By 'coordination' we mean that the radar emissions satisfy Pareto optimality with respect to multi-objective optimization over each radar's utility. This paper provides a novel multi-objective inverse reinforcement learning approach which allows for both detection of such Pareto optimal ('coordinating') behavior and subsequent reconstruction of each radar's utility function, given a finite dataset of radar network emissions. The method for accomplishing this is derived from the micro-economic setting of Revealed Preferences, and also applies to more general problems of inverse detection and learning of multi-objective optimizing systems.

A cognitive radar is a constrained utility maximizer that adapts its sensing mode in response to a changing environment. If an adversary can estimate the utility function of a cognitive radar, it can determine the radar's sensing strategy and mitigate the radar performance via electronic countermeasures (ECM). This paper discusses how a cognitive radar can {\em hide} its strategy from an adversary that detects cognition. The radar does so by transmitting purposefully designed sub-optimal responses to spoof the adversary's Neyman-Pearson detector. We provide theoretical guarantees by ensuring the Type-I error probability of the adversary's detector exceeds a pre-defined level for a specified tolerance on the radar's performance loss. We illustrate our cognition masking scheme via numerical examples involving waveform adaptation and beam allocation. We show that small purposeful deviations from the optimal strategy of the radar confuse the adversary by significant amounts, thereby masking the radar's cognition. Our approach uses novel ideas from revealed preference in microeconomics and adversarial inverse reinforcement learning. Our proposed algorithms provide a principled approach for system-level electronic counter-countermeasures (ECCM) to mask the radar's cognition, i.e., hide the radar's strategy from an adversary. We also provide performance bounds for our cognition masking scheme when the adversary has misspecified measurements of the radar's response.

Suppose $L$ simultaneous independent stochastic systems generate observations, where the observations from each system depend on the underlying parameter of that system. The observations are unlabeled (anonymized), in the sense that an analyst does not know which observation came from which stochastic system. How can the analyst estimate the underlying parameters of the $L$ systems? Since the anonymized observations at each time are an unordered set of L measurements (rather than a vector), classical stochastic gradient algorithms cannot be directly used. By using symmetric polynomials, we formulate a symmetric measurement equation that maps the observation set to a unique vector. By exploiting that fact that the algebraic ring of multi-variable polynomials is a unique factorization domain over the ring of one-variable polynomials, we construct an adaptive filtering algorithm that yields a statistically consistent estimate of the underlying parameters. We analyze the asymptotic covariance of these estimates to quantify the effect of anonymization. Finally, we characterize the anonymity of the observations in terms of the error probability of the maximum aposteriori Bayesian estimator. Using Blackwell dominance of mean preserving spreads, we construct a partial ordering of the noise densities which relates the anonymity of the observations to the asymptotic covariance of the adaptive filtering algorithm.

In mathematical psychology, recent models for human decision-making use Quantum Decision Theory to capture important human-centric features such as order effects and violation of the sure-thing principle (total probability law). We construct and analyze a human-sensor system where a quickest detector aims to detect a change in an underlying state by observing human decisions that are influenced by the state. Apart from providing an analytical framework for such human-sensor systems, we also analyze the structure of the quickest detection policy. We show that the quickest detection policy has a single threshold and the optimal cost incurred is lower bounded by that of the classical quickest detector. This indicates that intermediate human decisions strictly hinder detection performance. We also analyze the sensitivity of the quickest detection cost with respect to the quantum decision parameters of the human decision maker, revealing that the performance is robust to inaccurate knowledge of the decision-making process. Numerical results are provided which suggest that observing the decisions of more rational decision makers will improve the quickest detection performance. Finally, we illustrate a numerical implementation of this quickest detector in the context of the Prisoner's Dilemma problem, in which it has been observed that Quantum Decision Theory can uniquely model empirically tested violations of the sure-thing principle.

This paper considers the problem of estimating exposure to information in a social network. Given a piece of information (e.g., a URL of a news article on Facebook, a hashtag on Twitter), our aim is to find the fraction of people on the network who have been exposed to it. The exact value of exposure to a piece of information is determined by two features: the structure of the underlying social network and the set of people who shared the piece of information. Often, both features are not publicly available (i.e., access to the two features is limited only to the internal administrators of the platform) and difficult to be estimated from data. As a solution, we propose two methods to estimate the exposure to a piece of information in an unbiased manner: a vanilla method which is based on sampling the network uniformly and a method which non-uniformly samples the network motivated by the Friendship Paradox. We provide theoretical results which characterize the conditions (in terms of properties of the network and the piece of information) under which one method outperforms the other. Further, we outline extensions of the proposed methods to dynamic information cascades (where the exposure needs to be tracked in real-time). We demonstrate the practical feasibility of the proposed methods via experiments on multiple synthetic and real-world datasets.

In mathematical psychology, decision makers are modeled using the Lindbladian equations from quantum mechanics to capture important human-centric features such as order effects and violation of the sure thing principle. We consider human-machine interaction involving a quantum decision maker (human) and a controller (machine). Given a sequence of human decisions over time, how can the controller dynamically provide input messages to adapt these decisions so as to converge to a specific decision? We show via novel stochastic Lyapunov arguments how the Lindbladian dynamics of the quantum decision maker can be controlled to converge to a specific decision asymptotically. Our methodology yields a useful mathematical framework for human-sensor decision making. The stochastic Lyapunov results are also of independent interest as they generalize recent results in the literature., Comment: arXiv admin note: substantial text overlap with arXiv:2204.00059

Inverse reinforcement learning (IRL) deals with estimating an agent's utility function from its actions. In this paper, we consider how an agent can hide its strategy and mitigate an adversarial IRL attack; we call this inverse IRL (I-IRL). How should the decision maker choose its response to ensure a poor reconstruction of its strategy by an adversary performing IRL to estimate the agent's strategy? This paper comprises four results: First, we present an adversarial IRL algorithm that estimates the agent's strategy while controlling the agent's utility function. Our second result for I-IRL result spoofs the IRL algorithm used by the adversary. Our I-IRL results are based on revealed preference theory in micro-economics. The key idea is for the agent to deliberately choose sub-optimal responses that sufficiently masks its true strategy. Third, we give a sample complexity result for our main I-IRL result when the agent has noisy estimates of the adversary specified utility function. Finally, we illustrate our I-IRL scheme in a radar problem where a meta-cognitive radar is trying to mitigate an adversarial target.

This paper considers meta-cognitive radars in an adversarial setting. A cognitive radar optimally adapts its waveform (response) in response to maneuvers (probes) of a possibly adversarial moving target. A meta-cognitive radar is aware of the adversarial nature of the target and seeks to mitigate the adversarial target. How should the meta-cognitive radar choose its responses to sufficiently confuse the adversary trying to estimate the radar's utility function? This paper abstracts the radar's meta-cognition problem in terms of the spectra (eigenvalues) of the state and observation noise covariance matrices, and embeds the algebraic Riccati equation into an economics-based utility maximization setup. This adversarial target is an inverse reinforcement learner. By observing a noisy sequence of radar's responses (waveforms), the adversarial target uses a statistical hypothesis test to detect if the radar is a utility maximizer. In turn, the meta-cognitive radar deliberately chooses sub-optimal responses that increasing its Type-I error probability of the adversary's detector. We call this counter-adversarial step taken by the meta-cognitive radar as inverse inverse reinforcement learning (I-IRL). We illustrate the meta-cognition results of this paper via simple numerical examples. Our approach for meta-cognition in this paper is based on revealed preference theory in micro-economics and inspired by results in differential privacy and adversarial obfuscation in machine learning.

In mathematical psychology, decision makers are modeled using the Lindbladian equations from quantum mechanics to capture important human-centric features such as order effects and violation of the sure thing principle. We consider human-machine interaction involving a quantum decision maker (human) and a controller (machine). Given a sequence of human decisions over time, how can the controller dynamically provide input messages to adapt these decisions so as to converge to a specific decision? We show via novel stochastic Lyapunov arguments how the Lindbladian dynamics of the quantum decision maker can be controlled to converge to a specific decision asymptotically. Our methodology yields a useful mathematical framework for human-sensor decision making. The stochastic Lyapunov results are also of independent interest as they generalize recent results in the literature.

The paper constructs a multi-variate Hawkes process model of Bitcoin block arrivals and price jumps. Hawkes processes are selfexciting point processes that can capture the self- and cross-excitation effects of block mining and Bitcoin price volatility. We use publicly available blockchain datasets to estimate the model parameters via maximum likelihood estimation. The results show that Bitcoin price volatility boost block mining rate and Bitcoin investment return demonstrates mean reversion. Quantile-Quantile plots show that the proposed Hawkes process model is a better fit to the blockchain datasets than a Poisson process model.

Structural results impose sufficient conditions on the model parameters of a Markov decision process (MDP) so that the optimal policy is an increasing function of the underlying state. The classical assumptions for MDP structural results require supermodularity of the rewards and transition probabilities. However, supermodularity does not hold in many applications. This paper uses a sufficient condition for interval dominance (called I) proposed in the microeconomics literature, to obtain structural results for MDPs under more general conditions. We present several MDP examples where supermodularity does not hold, yet I holds, and so the optimal policy is monotone; these include sigmoidal rewards (arising in prospect theory for human decision making), bi-diagonal and perturbed bi-diagonal transition matrices (in optimal allocation problems). We also consider MDPs with TP3 transition matrices and concave value functions. Finally, reinforcement learning algorithms that exploit the differential sparse structure of the optimal monotone policy are discussed.

Tangle is a distributed ledger technology that stores data as a directed acyclic graph (DAG). Unlike blockchain, Tangle does not require dedicated miners for its operation; this makes Tangle suitable for Internet of Things (IoT) applications. Distributed ledgers have a built-in transaction rate control mechanism to prevent congestion and spamming; this is typically achieved by increasing or decreasing the proof of work (PoW) difficulty level based on the number of users. Unfortunately, this simplistic mechanism gives an unfair advantage to users with high computing power. This paper proposes a principal-agent problem (PAP) framework from microeconomics to control the transaction rate in Tangle. With users as agents and the transaction rate controller as the principal, we design a truth-telling mechanism to assign PoW difficulty levels to agents as a function of their computing power. The solution of the PAP is achieved by compensating a higher PoW difficulty level with a larger weight/reputation for the transaction. The mechanism has two benefits, (1) the security of Tangle is increased as agents are incentivized to perform difficult PoW, and (2) the rate of new transactions is moderated in Tangle. The solution of PAP is obtained by solving a mixed-integer optimization problem. We show that the optimal solution of the PAP increases with the computing power of agents. The structural results reduce the search space of the mixed-integer program and enable efficient computation of the optimal mechanism. Finally, via numerical examples, we illustrate the transaction rate control mechanism and study its impact on the dynamics of Tangle.

The paper develops a blockchain protocol for a social media network (BE-SMN) to mitigate the spread of misinformation. BE-SMN is derived based on the information transmission-time distribution by modeling the misinformation transmission as double-spend attacks on blockchain. The misinformation distribution is then incorporated into the SIR (Susceptible, Infectious, or Recovered) model, which substitutes the single rate parameter in the traditional SIR model. Then, on a multi-community network, we study the propagation of misinformation numerically and show that the proposed blockchain enabled social media network outperforms the baseline network in flattening the curve of the infected population.

We study how a cognitive radar can mask (hide) its cognitive ability from an adversarial jamming device. Specifically, if the radar optimally adapts its waveform based on adversarial target maneuvers (probes), how should the radar choose its waveform parameters (response) so that its utility function cannot be recovered by the adversary. This paper abstracts the radar's cognition masking problem in terms of the spectra (eigenvalues) of the state and observation noise covariance matrices, and embeds the algebraic Riccati equation into an economics-based utility maximization setup. Given an observed sequence of radar responses, the adversary tests for utility maximization behavior of the radar and estimates its utility function that rationalizes the radar's responses. In turn, the radar deliberately chooses sub-optimal responses so that its utility function almost fails the utility maximization test, and hence, its cognitive ability is masked from the adversary. We illustrate the performance of our cognition masking scheme via simple numerical examples. Our approach in this paper is based on revealed preference theory in microeconomics for identifying rationality.

This paper studies detecting anomalous edges in directed graphs that model social networks. We exploit edge exchangeability as a criterion for distinguishing anomalous edges from normal edges. Then we present an anomaly detector based on conformal prediction theory; this detector has a guaranteed upper bound for false positive rate. In numerical experiments, we show that the proposed algorithm achieves superior performance to baseline methods.

Electronic countermeasures (ECM) against a radar are actions taken by an adversarial jammer to mitigate effective utilization of the electromagnetic spectrum by the radar. On the other hand, electronic counter-countermeasures (ECCM) are actions taken by the radar to mitigate the impact of electronic countermeasures (ECM) so that the radar can continue to operate effectively. The main idea of this paper is to show that ECCM involving a radar and a jammer can be formulated as a principal-agent problem (PAP) - a problem widely studied in microeconomics. With the radar as the principal and the jammer as the agent, we design a PAP to optimize the radar's ECCM strategy in the presence of a jammer. The radar seeks to optimally trade-off signal-to-noise ratio (SNR) of the target measurement with the measurement cost: cost for generating radiation power for the pulse to probe the target. We show that for a suitable choice of utility functions, PAP is a convex optimization problem. Further, we analyze the structure of the PAP and provide sufficient conditions under which the optimal solution is an increasing function of the jamming power observed by the radar; this enables computation of the radar's optimal ECCM within the class of increasing affine functions at a low computation cost. Finally, we illustrate the PAP formulation of the radar's ECCM problem via numerical simulations. We also use simulations to study a radar's ECCM problem wherein the radar and the jammer have mismatched information.

This paper studies controlling segregation in social networks via exogenous incentives. We construct an edge formation game on a directed graph. A user (node) chooses the probability with which it forms an inter- or intra- community edge based on a utility function that reflects the tradeoff between homophily (preference to connect with individuals that belong to the same group) and the preference to obtain an exogenous incentive. Decisions made by the users to connect with each other determine the evolution of the social network. We explore an algorithmic recommendation mechanism where the exogenous incentive in the utility function is based on weak ties which incentivizes users to connect across communities and mitigates the segregation. This setting leads to a submodular game with a unique Nash equilibrium. In numerical simulations, we explore how the proposed model can be useful in controlling segregation and echo chambers in social networks under various settings.

This paper unifies two key results from economic theory, namely, revealed rational inattention and classical revealed preference. Revealed rational inattention tests for rationality of information acquisition for Bayesian decision makers. On the other hand, classical revealed preference tests for utility maximization under known budget constraints. Our first result is an equivalence result - we unify revealed rational inattention and revealed preference through an equivalence map over decision parameters and partial order for payoff monotonicity over the decision space in both setups. Second, we exploit the unification result computationally to extend robustness measures for goodness-of-fit of revealed preference tests in the literature to revealed rational inattention. This extension facilitates quantifying how well a Bayesian decision maker's actions satisfy rational inattention. Finally, we illustrate the significance of the unification result on a real-world YouTube dataset comprising thumbnail, title and user engagement metadata from approximately 140,000 videos. We compute the Bayesian analog of robustness measures from revealed preference literature on YouTube metadata features extracted from a deep auto-encoder, i.e., a deep neural network that learns low-dimensional features of the metadata. The computed robustness values show that YouTube user engagement fits the rational inattention model remarkably well. All our numerical experiments are completely reproducible.

How can non-communicating agents learn to share congested resources efficiently? This is a challenging task when the agents can access the same resource simultaneously (in contrast to multi-agent multi-armed bandit problems) and the resource valuations differ among agents. We present a fully distributed algorithm for learning to share in congested environments and prove that the agents' regret with respect to the optimal allocation is poly-logarithmic in the time horizon. Performance in the non-asymptotic regime is illustrated in numerical simulations. The distributed algorithm has applications in cloud computing and spectrum sharing. Keywords: Distributed learning, congestion games, poly-logarithmic regret.

Preferential attachment, homophily and, their consequences such as the glass ceiling effect have been well-studied in the context of undirected networks. However, the lack of an intuitive, theoretically tractable model of a directed, bi-populated~(i.e.,~containing two groups) network with variable levels of preferential attachment, homophily and growth dynamics~(e.g.,~the rate at which new nodes join, whether the new nodes mostly follow existing nodes or the existing nodes follow them, etc.) has largely prevented such consequences from being explored in the context of directed networks, where they more naturally occur due to the asymmetry of links. To this end, we present a rigorous theoretical analysis of the \emph{Directed Mixed Preferential Attachment} model and, use it to analyze the glass ceiling effect in directed networks. More specifically, we derive the closed-form expressions for the power-law exponents of the in- and out- degree distributions of each group~(minority and majority) and, compare them with each other to obtain insights. In particular, our results yield answers to questions such as: \emph{when does the minority group have a heavier out-degree (or in-degree) distribution compared to the majority group? what effect does frequent addition of edges between existing nodes have on the in- and out- degree distributions of the majority and minority groups?}. Such insights shed light on the interplay between the structure~(i.e., the in- and out- degree distributions of the two groups) and dynamics~(characterized collectively by the homophily, preferential attachment, group sizes and growth dynamics) of various real-world networks. Finally, we utilize the obtained analytical results to characterize the conditions under which the glass ceiling effect emerge in a directed network. Our analytical results are supported by detailed numerical results.

Structural inequalities persist in society, conferring systematic advantages to some people at the expense of others, for example, by giving them substantially more influence and opportunities. Using bibliometric data about authors of scientific publications, we identify two types of structural inequalities in scientific citations. First, female authors, who represent a minority of researchers, receive less recognition for their work (through citations) relative to male authors; second, authors affiliated with top-ranked institutions, who are also a minority, receive substantially more recognition compared to other authors. We present a model for the growth of directed citation networks and show that citations disparities arise from individual preferences to cite authors from the same group (homophily), highly cited or active authors (preferential attachment), as well as the size of the group and how frequently new authors join. We analyze the model and show that its predictions align well with real-world observations. Our theoretical and empirical analysis also suggests potential strategies to mitigate structural inequalities in science. In particular, we find that merely increasing the minority group size does little to narrow the disparities. Instead, reducing the homophily of each group, frequently adding new authors to a research field while providing them an accessible platform among existing, established authors, together with balanced group sizes can have the largest impact on reducing inequality. Our work highlights additional complexities of mitigating structural disparities stemming from asymmetric relations (e.g., directed citations) compared to symmetric relations (e.g., collaborations).

Are deep convolutional neural networks (CNNs) for image classification explainable by utility maximization with information acquisition costs? We demonstrate that deep CNNs behave equivalently (in terms of necessary and sufficient conditions) to rationally inattentive utility maximizers, a generative model used extensively in economics for human decision making. Our claim is based by extensive experiments on 200 deep CNNs from 5 popular architectures. The parameters of our interpretable model are computed efficiently via convex feasibility algorithms. As an application, we show that our economics-based interpretable model can predict the classification performance of deep CNNs trained with arbitrary parameters with accuracy exceeding 94% . This eliminates the need to re-train the deep CNNs for image classification. The theoretical foundation of our approach lies in Bayesian revealed preference studied in micro-economics. All our results are on GitHub and completely reproducible.

This paper deals with the modeling and estimation of the sociological phenomena called echo chambers and segregation in social networks. Specifically, we present a novel community-based graph model that represents the emergence of segregated echo chambers as a Markov bridge process. A Markov bridge is a one-dimensional Markov random field that facilitates modeling the formation and disassociation of communities at deterministic times which is important in social networks with known timed events. We justify the proposed model with six real world examples and examine its performance on a recent Twitter dataset. We provide model parameter estimation algorithm based on maximum likelihood and, a Bayesian filtering algorithm for recursively estimating the level of segregation using noisy samples obtained from the network. Numerical results indicate that the proposed filtering algorithm outperforms the conventional hidden Markov modeling in terms of the mean-squared error. The proposed filtering method is useful in computational social science where data-driven estimation of the level of segregation from noisy data is required.

It is well known that adding any skew symmetric matrix to the gradient of Langevin dynamics algorithm results in a non-reversible diffusion with improved convergence rate. This paper presents a gradient algorithm to adaptively optimize the choice of the skew symmetric matrix. The resulting algorithm involves a non-reversible diffusion algorithm cross coupled with a stochastic gradient algorithm that adapts the skew symmetric matrix. The algorithm uses the same data as the classical Langevin algorithm. A weak convergence proof is given for the optimality of the choice of the skew symmetric matrix. The improved convergence rate of the algorithm is illustrated numerically in Bayesian learning and tracking examples.