Transformer-based forecasters have recently become state-of-the-art in time series predictions over linear forecasters. However, this promotion comes at the loss of the innate interpretability of linear forecasters. Real life examples of time series prediction often involve critical data, and interpretability is crucial for continuous improvement and accountability when predicting. This paper introduces a novel approach that extends existing transformer-specific interpretability methodologies to bridge this gap. This paper exploits relevance calculation that are employed in adjacent transformer-based architecture in image and natural language processing. This method is then applied to the inverted transformer architecture, one of the latest and best performing transformer-based forecaster. The method produces a relevance map linking the relationship drawn between features by the model. The relevance map gives feedback on the key variables that influence the model’s decision the most, further enhancing the human understanding of the dataset. A key result of this study is a significant improvement in transformer based model explainability, without relying directly on visualizing the attention weights. Another key outcome is the potential adaptation of other transformer-specific tools to the inverted transformer.

This paper introduces a novel centralized multi-agent framework employing proximal policy optimization (PPO), a state-of-the-art reinforcement learning algorithm. Multiple subagents focus on different aspects of the environment, and the actions suggested by the subagents are used to augment the environment space for a superagent that encompasses the subagents' different approaches. A key feature of this model is the built-in attention module that balances the weight attributed to the environment variables and the suggested actions. This architecture is designed to promote emergent behaviour and exploration, completing the standard training of a PPO agent. We study the efficiency and decision accuracy from the model across several MuJoCo scenarios, and perform an ablation study to demonstrate the influence of the attention mechanism. Our results indicate a tradeoff in performance linked to the dimensions of the action space and address the ideal use case for this framework.

Efficient spectrum use in wireless sensor networks through spatial reuse requires effective models of packet reception at the physical layer in the presence of interference. Despite recent progress in analytic and simulations research into worst-case behavior from interference effects, these efforts generally assume geometric path loss and isotropic transmission, assumptions which have not been borne out in experiments.

We present a framework for the maximum flow problem in wireless networks using the SINR interference model in combination with MIMO nodes. The performance ratio of our algorithm matches the best O(log n) approximation factor known for the pure flow problem, but avoids the impractical dependence on the ellipsoid method and features both simpler and more intuitive analysis. The objective of a maximum flow in wireless networks is to get as much information from a sender to a receiver using intermediate nodes in a multi-hop environment. The algorithm is based on an LP formulation of the flow problem, and handles gracefully all additional linear constraints. The set of constraints that can be included contains, among other, power limits, fairness between source-sink pairs, capacity limits and bounds, and the use of Multi-Input Multi-Output nodes for the source and the sink nodes, which are often the bottlenecks in a wireless flow.

No abstract available.

We consider the capacity problem (or, the single slot scheduling problem) in wireless networks. Our goal is to maximize the number of successful connections in arbitrary wirelessnetworks where a transmission is successful only if the signal-to-interference-plus-noise ratio at the receiver is greater than some threshold. We study a game theoretic approach towards capacity maximization introduced by Andrews and Dinitz (INFOCOM 2009) and Dinitz (INFOCOM 2010). We prove vastly improved bounds for the game theoretic algorithm. In doing so, we achieve the first distributed constant factor approximation algorithm for capacity maximization for the uniform power assignment. When compared to the optimum where links may use an arbitrary power assignment, we prove a O(log Δ) approximation, where Δ is the ratio between the largest and the smallest link in the network. This is an exponential improvement of the approximation factor compared to existing results for distributed algorithms. All our results work for links located in any metric space. In addition, we provide simulation studies clarifying the picture on distributed algorithms for capacity maximization.

The vertex cover problem is a classical NP-complete problem for which the best worst-case approximation ratio is $2-o(1)$. In this paper, we use a collection of simple graph transformations, each of which guarantees an approximation ratio of $\frac{3}{2}$, to find approximate vertex covers for a large collection of randomly generated graphs and test graphs from various sources. The graph reductions are extremely fast, and even though they by themselves are not guaranteed to find a vertex cover, we manage to find a $\frac{3}{2}$-approximate vertex cover for almost every single graph in our collection. The few graphs that we cannot solve have specific structure: they are triangle-free, with a minimum degree of at least 3, a lower bound of $\frac{n}{2}$ on the optimal vertex cover, and are unlikely to have a large bipartite subgraph.

We show how the on-line bounded-space bin cover problem can be modeled with a Markov chain. We then use this Markov chain formulation to derive an algorithm for the on-line bounded-space bin cover problem. Our algorithm is designed to perform well in a restrictive environment where it can utilize only very few open bins at each time. We analyze the performance of our algorithm and compare it to the Sum-of-Squares with Threshold algorithm. The experimental results show that our algorithm compares favorably with the Sum-of-Squares with Threshold algorithm, and the average waste incurred by our algorithm is very small even when it is forced to use only a handful of open bins.