arXiv subjects

Mohan Dantam

Publications and source records attributed to Mohan Dantam.

5 recordsLinked to original sources

arXiv2026

Mean-Payoff-Parity and Lifting Strategies from MDPs to 2-Player Stochastic Games

We consider the strategy complexity (i.e., memory and randomization) of optimal strategies in turn-based 2-player zero-sum stochastic games. Results in [Gimbert,Kelmendi:2023] show how to lift optimal memoryless strategies for shift-invariant inverse-submixing objectives from MDPs to 2-player stochastic games with an exponential increase in the number of memory modes. We show the corresponding lower bound, i.e., the extra exponential memory is required in general, even for randomized strategies. Moreover, we solve the strategy complexity of the well-studied mean-payoff-parity objective in 2-player stochastic games. This objective is also shift-invariant inverse-submixing, but easier than the worst case for this class. In MDPs, Maximizer has optimal memoryless randomized strategies, while optimal deterministic strategies require exponential memory. However, in stochastic games, optimal randomized strategies require, at least and at most, linear memory (equal to the number of even colors). Finally, we show that a different construction for lifting memoryless (resp. finite-memory) deterministic strategies from MDPs (resp. 1-player games) to 2-player games cannot be generalized even to memoryless randomized strategies. We construct a shift-invariant objective where Max and Min each have optimal memoryless randomized strategies in all MDPs, but optimal (randomized) Max strategies still require infinite memory in deterministic 2-player games.

cs.GT↗

arXiv2024

Finite-memory Strategies for Almost-sure Energy-MeanPayoff Objectives in MDPs

We consider finite-state Markov decision processes with the combined Energy-MeanPayoff objective. The controller tries to avoid running out of energy while simultaneously attaining a strictly positive mean payoff in a second dimension. We show that finite memory suffices for almost surely winning strategies for the Energy-MeanPayoff objective. This is in contrast to the closely related Energy-Parity objective, where almost surely winning strategies require infinite memory in general. We show that exponential memory is sufficient (even for deterministic strategies) and necessary (even for randomized strategies) for almost surely winning Energy-MeanPayoff. The upper bound holds even if the strictly positive mean payoff part of the objective is generalized to multidimensional strictly positive mean payoff. Finally, it is decidable in pseudo-polynomial time whether an almost surely winning strategy exists.

cs.GT↗

arXiv2023

Inference and Interference: The Role of Clipping, Pruning and Loss Landscapes in Differentially Private Stochastic Gradient Descent

Differentially private stochastic gradient descent (DP-SGD) is known to have poorer training and test performance on large neural networks, compared to ordinary stochastic gradient descent (SGD). In this paper, we perform a detailed study and comparison of the two processes and unveil several new insights. By comparing the behavior of the two processes separately in early and late epochs, we find that while DP-SGD makes slower progress in early stages, it is the behavior in the later stages that determines the end result. This separate analysis of the clipping and noise addition steps of DP-SGD shows that while noise introduces errors to the process, gradient descent can recover from these errors when it is not clipped, and clipping appears to have a larger impact than noise. These effects are amplified in higher dimensions (large neural networks), where the loss basin occupies a lower dimensional space. We argue theoretically and using extensive experiments that magnitude pruning can be a suitable dimension reduction technique in this regard, and find that heavy pruning can improve the test accuracy of DPSGD.

cs.LG↗

arXiv2023

Approximating the Value of Energy-Parity Objectives in Simple Stochastic Games

We consider simple stochastic games $\mathcal G$ with energy-parity objectives, a combination of quantitative rewards with a qualitative parity condition. The Maximizer tries to avoid running out of energy while simultaneously satisfying a parity condition. We present an algorithm to approximate the value of a given configuration in 2-NEXPTIME. Moreover, $\varepsilon$-optimal strategies for either player require at most $O(2EXP(|{\mathcal G}|)\cdot\log(\frac{1}{\varepsilon}))$ memory modes.

cs.GT↗

arXiv2020

On the Decidability of Reachability in Continuous Time Linear Time-Invariant Systems

We consider the decidability of state-to-state reachability in linear time-invariant control systems over continuous time. We analyse this problem with respect to the allowable control sets, which are assumed to be the image under a linear map of the unit hypercube. This naturally models bounded (sometimes called saturated) controls. Decidability of the version of the reachability problem in which control sets are affine subspaces of $\mathbb{R}^n$ is a fundamental result in control theory. Our first result is decidability in two dimensions ($n=2$) if the matrix $A$ satisfies some spectral conditions, and conditional decidablility in general. If the transformation matrix $A$ is diagonal with rational entries (or rational multiples of the same algebraic number) then the reachability problem is decidable. If the transformation matrix $A$ only has real eigenvalues, the reachability problem is conditionally decidable. The time-bounded reachability problem is conditionally decidable, and unconditionally decidable in two dimensions. Some of our decidability results are conditional in that they rely on the decidability of certain mathematical theories, namely the theory of the reals with exponential ($\mathfrak{R}_{\exp}$) and with bounded sine ($\mathfrak{R}_{\exp,\sin}$). We also obtain a hardness result for a mild generalization of the problem where the target is simple set (hypercube of dimension $n-1$ or hyperplane) instead of a point, and the control set is a convex bounded polytope. In this case, we show that the problem is at least as hard as the \emph{Continuous Positivity problem} or the \emph{Nontangential Continuous Positivity problem}.

math.OC↗