Searcharxiv⌕ Search

arXiv subjects

Atilla Eryilmaz

Publications and source records attributed to Atilla Eryilmaz.

At least 19 recordsLinked to original sources

Finite-Memory Control of POMDPs: Fundamental Limits and Efficient Design

Optimal control of partially observable Markov decision processes (POMDPs), in general, requires a complete history. The belief state, a distribution over hidden states, summarizes this history and provides the classical sufficient statistic to control POMDPs. However, the continuous nature of the belief space generally prevents a finite-memory controller from maintaining it exactly. This raises the question we address in this work: what must the available finite memory preserve to achieve near-optimal control performance? We study this question through an information-theoretic converse that gives a fundamental lower bound on the loss of any finite-memory controller. The analysis introduces the concept of a decision witness that characterizes belief states whose confusion would lead to unavoidable loss. Motivated by this bound, we propose RECAP, a constructive quantization method for designing representative memory states. We prove an upper bound on the optimality gap of any RECAP codebook and use the components of this bound together with the converse analysis to define the codebook objective. Experiments on 11 POMDP benchmarks demonstrate that RECAP improves over finite-memory representation baselines under the same memory budget.

cs.LG↗

Sharp Convergence and Sample Complexity of Policy Mirror Descent for Average-Reward MDPs

Policy mirror descent (PMD) has a mature finite-time theory in discounted Markov decision processes (MDPs), but less is known in the average-reward setting, a more natural objective for many control applications. We give a finite-time, finite-sample analysis of PMD in ergodic average-reward MDPs built around a single master recursion that governs convergence for any critic, without external regularization. Its specializations yield linear rates for exact, inexact-tabular, and linear function approximation (LFA) updates. For exact PMD, the rate is tight in the worst case, and convergence can become superlinear. We complement these convergence results with critics whose end-to-end sample complexities improve the prior best policy-gradient dependence from $|S||A|\,t_{\mathrm{mix}}^3$ to $|S|\,t_{\mathrm{mix}}^3+|S||A|\,t_{\mathrm{mix}}^2$ in the tabular setting and from $|S||A|\,t_{\mathrm{mix}}^5$ to $d\,t_{\mathrm{mix}}^3$ under LFA. Matching information-theoretic lower bounds establish that both critics are optimal up to logarithmic factors.

cs.LG↗

Clock-Free Stochastic Scheduling for Real-Time Synchronized Content Delivery

Real-time synchronized content delivery is critical for applications such as cloud gaming, collaborative mixed reality, and live streaming, where a server must ensure that a group of users perceives the same content simultaneously for each delivery. Due to heterogeneous and unpredictable network delays, ensuring synchronization is highly challenging. We propose a clock-free scheduling framework that introduces latency compensation to control the server's delivery time for each user, aiming to maximize synchronization performance (i.e., minimizing the gap between the maximum and minimum latencies) while maintaining a low overall latency cost. Unlike prior work requiring synchronized clocks or client-side timestamps, our approach uses only acknowledgment feedback, enabling flexible deployment across diverse environments. We formulate the problem as a convex stochastic optimization with a non-smooth objective. To solve the problem, we employ the stochastic proximal point (SPP) algorithm that updates the delay compensation, using only the previous latency samples, rather than the latency distribution. By exploiting the subproblem structure, we apply ADMM to enable efficient updates with per-iteration complexity of at most $\mathcal{O}(N\log N)$, where $N$ is the number of users. We establish a convergence rate of $\mathcal{O}(1/\sqrt{K})$, where $K$ is the number of content deliveries for transmission. The framework is further extended to handle transmission failures, where packet loss results in successful delivery to only a subset of users. Finally, simulations using both synthetic and real-world latency traces demonstrate the consistent superiority of our algorithm against baseline methods.

cs.NI↗

PMF-CL: Pareto-Minimal-Forgetting Continual Learner for Conflicting Tasks

In the literature, many continual learning (CL) algorithms have been proposed to address the issue of catastrophic forgetting in ML models (i.e., learning new tasks leads to the loss of performance on previously learned tasks). Although all CL approaches use some form of memory to retain information about past tasks, a grounded understanding of what information needs to be stored to minimize catastrophic forgetting remains elusive. Recently, it has been recognized that under the strong assumption of the existence of a common global minimizer over all tasks, catastrophic forgetting can be completely avoided. However, in practice, tasks rarely have a common global minimizer, and a certain amount of forgetting is inevitable. In this paper, we propose a foundational reframing of CL as balancing conflicting tasks in hindsight from a multi-task learning (MTL) perspective. The approach is based on finding Pareto-optimal solutions, i.e., the solutions which, by definition, minimally forget the previous tasks in the Pareto sense. We derive a Pareto-minimal-forgetting CL (PMF-CL) algorithm for linear and basis-function regression. Our algorithm naturally extends to loss functions with quadratic upper bounds around their minimizers, despite consuming only a static memory footprint of $O(d^2)$ for $d$ model parameters, independent of the number of sequentially occurring tasks $T$. In the quadratic-loss setting, we obtain exact hindsight Pareto-optimal guarantees on forgetting. In the quadratic-upper-bound setting, forgetting is provably bounded, with the tightness of the forgetting bound being directly determined by the tightness of the quadratic bound on the loss function. Our numerical results validate our theoretical claims of convergence to the Pareto-optimal solution, and demonstrate that we consistently out-perform prior empirical methods while occupying lesser memory.

cs.LG↗

A Comparative Analysis of Attention versus State-Space Models for In-Context Learning

Transformers and state-space models (SSMs) are two prominent sequential learning architectures, yet their comparison remains largely empirical and existing theoretical analyses are typically task-specific or architecturally restricted. In this paper, we develop belief geometry, a unified analytical framework for comparing the representational capabilities of broad classes of attention and SSMs. Starting from a generalized formulation of in-context linear regression and using cumulative Bayes regret as our measure, we abstract three capabilities required by many sequential learning problems in our belief geometry: evidence assembly, belief maintenance, and addressing. We then study three cases of our formulation that isolate these capabilities and yield sharp architectural lessons: For belief maintenance, SSMs attain the optimal regret over stationary aggregation kernels; for positional assembly, SSMs have a memory advantage; and for content addressing, softmax attention has an exponential width advantage over sigmoid-selective SSMs. Experiments with LLaMA-type Transformers and Mamba-2 show that these architectural insights extend beyond our analytically tractable classes and linear-regression testbed.

cs.LG↗

Communication-Aware Model Distributed Inference via Latent Representation Compression

We study optimization of distributed model inference over resource-constrained edge resources. We propose a framework that optimizes the trade-off between model accuracy and communication costs by controlling latent representation compression to meet strict Quality of Service (QoS) throughput targets. For settings with known channel state information (CSI), we derive a closed-form optimal solution for single tasks and reduce the multi-task problem to a convex optimization program characterized by a per-link water-filling strategy. We extend these to handle unpredictable environments via a stochastic dual descent algorithm that relies only on causal channel estimates. We provide Lyapunov-based proofs demonstrating that our approach strictly satisfies long-term delay constraints while achieving a bounded optimality gap. Our results offer a robust, scalable blueprint for maximizing the performance of pipelined AI tasks in dynamic, resource-constrained distributed systems. We verify the effectiveness of our proposed framework through simulations and experiments with real edge devices.

cs.DC↗

Finite-Time Analysis of Gradient Descent for Shallow Transformers

Understanding why Transformers perform so well remains challenging due to their non-convex optimization landscape. In this work, we analyze a shallow Transformer with $m$ independent heads trained by projected gradient descent in the kernel regime. Our analysis reveals two main findings: (i) the width required for nonasymptotic guarantees scales only logarithmically with the sample size $n$, and (ii) the optimization error is independent of the sequence length $T$. This contrasts sharply with recurrent architectures, where the optimization error can grow exponentially with $T$. The trade-off is memory: to keep the full context, the Transformer's memory requirement grows with the sequence length. We validate our theoretical results numerically in a teacher-student setting and compare Transformers with recurrent architectures on an autoregressive task.

cs.LG↗

An LP-based Sampling Policy for Multi-Armed Bandits with Side-Observations and Stochastic Availability

We study the stochastic multi-armed bandit (MAB) problem where an underlying network structure enables side-observations across related actions. We use a bipartite graph to link actions to a set of unknowns, such that selecting an action reveals observations for all the unknowns it is connected to. While previous works rely on the assumption that all actions are permanently accessible, we investigate the more practical setting of stochastic availability, where the set of feasible actions (the "activation set") varies dynamically in each round. This framework models real-world systems with both structural dependencies and volatility, such as social networks where users provide side-information about their peers' preferences, yet are not always online to be queried. To address this challenge, we propose UCB-LP-A, a novel policy that leverages a Linear Programming (LP) approach to optimize exploration-exploitation trade-offs under stochastic availability. Unlike standard network bandit algorithms that assume constant access, UCB-LP-A computes an optimal sampling distribution over the realizable activation sets, ensuring that the necessary observations are gathered using only the currently active arms. We derive a theoretical upper bound on the regret of our policy, characterizing the impact of both the network structure and the activation probabilities. Finally, we demonstrate through numerical simulations that UCB-LP-A significantly outperforms existing heuristics that ignore either the side-information or the availability constraints.

cs.LG↗

Recurrent Natural Policy Gradient for POMDPs

Solving partially observable Markov decision processes (POMDPs) remains a fundamental challenge in reinforcement learning (RL), primarily due to the curse of dimensionality induced by the non-stationarity of optimal policies. In this work, we study a natural actor-critic (NAC) algorithm that integrates recurrent neural network (RNN) architectures into a natural policy gradient (NPG) method and a temporal difference (TD) learning method. This framework leverages the representational capacity of RNNs to address non-stationarity in RL to solve POMDPs while retaining the statistical and computational efficiency of natural gradient methods in RL. We provide non-asymptotic theoretical guarantees for this method, including bounds on sample and iteration complexity to achieve global optimality up to function approximation. Additionally, we characterize pathological cases that stem from long-term dependencies, thereby explaining limitations of RNN-based policy optimization for POMDPs.

math.OC↗

Optimal Parallel Scheduling under Concave Speedup Functions

Efficient scheduling of parallel computation resources across multiple jobs is a fundamental problem in modern cloud/edge computing systems for many AI-based applications. Allocating more resources to a job accelerates its completion, but with diminishing returns. Prior work (heSRPT) solved this problem only for some specific speedup functions with an exponential form, providing a closed-form solution. However, the general case with arbitrary concave speedup functions -- which more accurately capture real-world workloads -- has remained open. In this paper, we solve this open problem by developing optimal scheduling algorithms for parallel jobs under general concave speedup functions. We first discover a fundamental and broadly-applicable rule for optimal parallel scheduling, namely the Consistent Derivative Ratio (CDR) Rule, which states that the ratio of the derivatives of the speedup functions across active jobs remains constant over time. To efficiently compute the optimal allocations that satisfy the CDR Rule, we propose the General Water-Filling (GWF) method, a more general version of classical water-filling in wireless communications. Combining these insights, we design the SmartFill Algorithm to solve the general scheduling problem. Unlike heSRPT, which always allocates resources to all active jobs, SmartFill selectively determines which jobs should receive resources and how much they should be allocated. For a broad class of so-called \emph{regular} speedup functions, SmartFill yields closed-form optimal solutions, while for non-regular functions it efficiently computes the optimum with low complexity. Numerical evaluations show that SmartFill can substantially outperform heSRPT across a wide range of concave speedup functions.

cs.DC↗

BeST -- A Novel Source Selection Metric for Transfer Learning

One of the most fundamental, and yet relatively less explored, goals in transfer learning is the efficient means of selecting top candidates from a large number of previously trained models (optimized for various "source" tasks) that would perform the best for a new "target" task with a limited amount of data. In this paper, we undertake this goal by developing a novel task-similarity metric (BeST) and an associated method that consistently performs well in identifying the most transferrable source(s) for a given task. In particular, our design employs an innovative quantization-level optimization procedure in the context of classification tasks that yields a measure of similarity between a source model and the given target data. The procedure uses a concept similar to early stopping (usually implemented to train deep neural networks (DNNs) to ensure generalization) to derive a function that approximates the transfer learning mapping without training. The advantage of our metric is that it can be quickly computed to identify the top candidate(s) for a given target task before a computationally intensive transfer operation (typically using DNNs) can be implemented between the selected source and the target task. As such, our metric can provide significant computational savings for transfer learning from a selection of a large number of possible source models. Through extensive experimental evaluations, we establish that our metric performs well over different datasets and varying numbers of data samples.

cs.LG↗

Convergence of Gradient Descent for Recurrent Neural Networks: A Nonasymptotic Analysis

We analyze recurrent neural networks with diagonal hidden-to-hidden weight matrices, trained with gradient descent in the supervised learning setting, and prove that gradient descent can achieve optimality \emph{without} massive overparameterization. Our in-depth nonasymptotic analysis (i) provides improved bounds on the network size $m$ in terms of the sequence length $T$, sample size $n$ and ambient dimension $d$, and (ii) identifies the significant impact of long-term dependencies in the dynamical system on the convergence and network width bounds characterized by a cutoff point that depends on the Lipschitz continuity of the activation function. Remarkably, this analysis reveals that an appropriately-initialized recurrent neural network trained with $n$ samples can achieve optimality with a network size $m$ that scales only logarithmically with $n$. This sharply contrasts with the prior works that require high-order polynomial dependency of $m$ on $n$ to establish strong regularity conditions. Our results are based on an explicit characterization of the class of dynamical systems that can be approximated and learned by recurrent neural networks via norm-constrained transportation mappings, and establishing local smoothness properties of the hidden state with respect to the learnable parameters.

cs.LG↗

Non-Federated Multi-Task Split Learning for Heterogeneous Sources

With the development of edge networks and mobile computing, the need to serve heterogeneous data sources at the network edge requires the design of new distributed machine learning mechanisms. As a prevalent approach, Federated Learning (FL) employs parameter-sharing and gradient-averaging between clients and a server. Despite its many favorable qualities, such as convergence and data-privacy guarantees, it is well-known that classic FL fails to address the challenge of data heterogeneity and computation heterogeneity across clients. Most existing works that aim to accommodate such sources of heterogeneity stay within the FL operation paradigm, with modifications to overcome the negative effect of heterogeneous data. In this work, as an alternative paradigm, we propose a Multi-Task Split Learning (MTSL) framework, which combines the advantages of Split Learning (SL) with the flexibility of distributed network architectures. In contrast to the FL counterpart, in this paradigm, heterogeneity is not an obstacle to overcome, but a useful property to take advantage of. As such, this work aims to introduce a new architecture and methodology to perform multi-task learning for heterogeneous data sources efficiently, with the hope of encouraging the community to further explore the potential advantages we reveal. To support this promise, we first show through theoretical analysis that MTSL can achieve fast convergence by tuning the learning rate of the server and clients. Then, we compare the performance of MTSL with existing multi-task FL methods numerically on several image classification datasets to show that MTSL has advantages over FL in training speed, communication cost, and robustness to heterogeneous data.

cs.LG↗

SwiftCache: Model-Based Learning for Dynamic Content Caching in CDNs

We introduce SwiftCache, a "fresh" learning-based caching framework designed for content distribution networks (CDNs) featuring distributed front-end local caches and a dynamic back-end database. Users prefer the most recent version of the dynamically updated content, while the local caches lack knowledge of item popularity and refresh rates. We first explore scenarios with requests arriving at a local cache following a Poisson process, whereby we prove that the optimal policy features a threshold-based structure with updates occurring solely at request arrivals. Leveraging these findings, SwiftCache is proposed as a model-based learning framework for dynamic content caching. The simulation demonstrates near-optimal cost for Poisson process arrivals and strong performance with limited cache sizes. For more general environments, we present a model-free Reinforcement Learning (RL) based caching policy without prior statistical assumptions. The model-based policy performs well compared to the model-free policy when the variance of interarrival times remains moderate. However, as the variance increases, RL slightly outperforms model-based learning at the cost of longer training times, and higher computational resource consumption. Model-based learning's adaptability to environmental changes without retraining positions it as a practical choice for dynamic network environments. Distributed edge caches can utilize this approach in a decentralized manner to effectively meet the evolving behaviors of users.

math.OC↗

Optimal Push and Pull-Based Edge Caching For Dynamic Content

We introduce a framework and optimal `fresh' caching for a content distribution network (CDN) comprising a front-end local cache and a back-end database. The data content is dynamically updated at a back-end database and end-users are interested in the most-recent version of that content. We formulate the average cost minimization problem that captures the system's cost due to the service of aging content as well as the regular cache update cost. We consider the cost minimization problem from two individual perspectives based on the available information to either side of the CDN: the back-end database perspective and the front-end local cache perspective. For the back-end database, the instantaneous version of content is observable but the exact demand is not. Caching decisions made by the back-end database are termed `push-based caching'. For the front-end local cache, the age of content version in the cache is not observable, yet the instantaneous demand is. Caching decisions made by the front-end local cache are termed `pull-based caching'. Our investigations reveal which type of information, updates, or demand dynamic, is of higher value towards achieving the minimum cost based on other network parameters including content popularity, update rate, and demand intensity.

math.OC↗

Optimal Edge Caching For Individualized Demand Dynamics

The ever-growing end user data demands, and the simultaneous reductions in memory costs are fueling edge-caching deployments. Caching at the edge is substantially different from that at the core and needs to take into account the nature of individual data demands. For example, an individual user may not be interested in requesting the same data item again, if it has recently requested it. Such individual dynamics are not apparent in the aggregated data requests at the core and have not been considered in popularity-driven caching designs for the core. Hence, these traditional caching policies could induce significant inefficiencies when applied at the edges. To address this issue, we develop new edge caching policies optimized for the individual demands that also leverage overhearing opportunities at the wireless edge. With the objective of maximizing the hit ratio, the proposed policies will actively evict the data items that are not likely to be requested in the near future, and strategically bring them back into the cache through overhearing when they are likely to be popular again. Both theoretical analysis and numerical simulations demonstrate that the proposed edge caching policies could outperform the popularity-driven policies that are optimal at the core.

cs.NI↗

Provably Robust Temporal Difference Learning for Heavy-Tailed Rewards

In a broad class of reinforcement learning applications, stochastic rewards have heavy-tailed distributions, which lead to infinite second-order moments for stochastic (semi)gradients in policy evaluation and direct policy optimization. In such instances, the existing RL methods may fail miserably due to frequent statistical outliers. In this work, we establish that temporal difference (TD) learning with a dynamic gradient clipping mechanism, and correspondingly operated natural actor-critic (NAC), can be provably robustified against heavy-tailed reward distributions. It is shown in the framework of linear function approximation that a favorable tradeoff between bias and variability of the stochastic gradients can be achieved with this dynamic gradient clipping mechanism. In particular, we prove that robust versions of TD learning achieve sample complexities of order $\mathcal{O}(\varepsilon^{-\frac{1}{p}})$ and $\mathcal{O}(\varepsilon^{-1-\frac{1}{p}})$ with and without the full-rank assumption on the feature matrix, respectively, under heavy-tailed rewards with finite moments of order $(1+p)$ for some $p\in(0,1]$, both in expectation and with high probability. We show that a robust variant of NAC based on Robust TD learning achieves $\tilde{\mathcal{O}}(\varepsilon^{-4-\frac{2}{p}})$ sample complexity. We corroborate our theoretical results with numerical experiments.

cs.LG↗

Age-Optimal Multi-Channel-Scheduling under Energy and Tolerance Constraints

We study the optimal scheduling problem where n source nodes attempt to transmit updates over L shared wireless on/off fading channels to optimize their age performance under energy and age-violation tolerance constraints. Specifically, we provide a generic formulation of age-optimization in the form of a constrained Markov Decision Processes (CMDP), and obtain the optimal scheduler as the solution of an associated Linear Programming problem. We investigate the characteristics of the optimal single-user multi-channel scheduler for the important special cases of average-age and violation-rate minimization. This leads to several key insights on the nature of the optimal allocation of the limited energy, where a usual threshold-based policy does not apply and will be useful in guiding scheduler designers. We then investigate the stability region of the optimal scheduler for the multi-user case. We also develop an online scheduler using Lyapunov-drift-minimization methods that do not require the knowledge of channel statistics. Our numerical studies compare the stability region of our online scheduler to the optimal scheduler to reveal that it performs closely with unknown channel statistics.

cs.IT↗