The widespread deployment of machine learning techniques in ubiquitous computing environments has sparked interests in exploiting the vast amount of data stored on mobile devices. To preserve data privacy, federated learning (FL) has been proposed to learn a shared model by performing distributed training locally on participating devices and aggregating the local models into a global one. Reinforcement learning (RL) can improve indoor localization by accounting for environmental dynamics, but has been trained on centralized data. An FL version of RL can help train a global localization model using data from different user clients whereas keeping data on device without centralization. We propose a personalized federated RL for indoor localization that addresses two major challenges. Due to the limited network connectivity of mobile devices, under the federated computing setting, it is impractical to aggregate updates from all clients in any learning iteration. Data gathered on different devices are heterogeneous, imposing difficulty in training high accuracy models. In our approach, each client performs RL to learn an action policy that can quickly search for a target based on its own data (e.g., personalized) and then a central server communicates with clients only for their model updates and learns a global model that is in the proximity of all client models (e.g., federated). Empirical evaluations demonstrate superior performance of the proposed approach in terms of localization accuracy and steadiness over existing methods. We further extend our approach to few-shot learning that can quickly position a new user with sparse annotated location data.

Learning feature interactions can be the key for multivariate predictive modeling. ReLU-activated neural networks create piecewise linear prediction models, and other nonlinear activation functions lead to models with only high-order feature interactions. Recent methods incorporate candidate polynomial terms of fixed orders into deep learning, which is subject to the issue of combinatorial explosion, or learn the orders that are difficult to adapt to different regions of the feature space. We propose a Polyhedron Attention Module (PAM) to create piecewise polynomial models where the input space is split into polyhedrons which define the different pieces and on each piece the hyperplanes that define the polyhedron boundary multiply to form the interactive terms, resulting in interactions of adaptive order to each piece. PAM is interpretable to identify important interactions in predicting a target. Theoretic analysis shows that PAM has stronger expression capability than ReLU-activated networks. Extensive experimental results demonstrate the superior classification performance of PAM on massive datasets of the click-through rate prediction and PAM can learn meaningful interaction effects in a medical problem.

We consider the problem of clustering graph nodes over large-scale dynamic graphs, such as citation networks, images and web networks, when graph updates such as node/edge insertions/deletions are observed distributively. We propose communication-efficient algorithms for two well-established communication models namely the message passing and the blackboard models. Given a graph with n nodes that is observed at s remote sites over time [1,t], the two proposed algorithms have communication costs Õ(ns) and Õ(n + s) (Õ hides a polylogarithmic factor), almost matching their lower bounds, Ω(ns) and Ω(n + s), respectively, in the message passing and the blackboard models. More importantly, we prove that at each time point in [1,t] our algorithms generate clustering quality nearly as good as that of centralizing all updates up to that time and then applying a standard centralized clustering algorithm. We conducted extensive experiments on both synthetic and real-life datasets which confirmed the communication efficiency of our approach over baseline algorithms while achieving comparable clustering results.

Molecular similarity search is a simple but powerful chemoinformatics tool to rapidly find molecules that are structurally similar to a known reference compound from a large molecular database. A variety of indexing structures had been developed to improve the performance of similarity search over the large compound database. However, those algorithms often require a large computational cost to build indices and process queries, especially for a large-scale molecular dataset. We study the problem of accelerating similarity search using high performance computing (HPC) and design general algorithms to speed up existing indexing algorithms. We first propose a parallel algorithm based on data chunking, working for all indexing algorithms for similarity search. We theoretically analyze its computation cost and relationships between the speedup and number of data chunks. We further propose a parallel query algorithm for all graph-based indexing algorithms to accelerate their query processing in HPC. Both of our algorithms consistently offer a greater speedup than the baseline algorithm(s) when evaluated with different datasets and parameter settings.

Phantom Go is a derivative of Go with imperfect information. It is challenging in AI field due to its great uncertainty of the hidden information and high game complexity inherited from Go. To deal with this imperfect information game with large game tree complexity, a general search framework named belief-state Monte Carlo tree search (BS-MCTS) is put forward in this paper. BS-MCTS incorporates belief-states into Monte Carlo Tree Search, where belief-state is a notation derived from philosophy to represent the probability that speculation is in accordance with reality. In BS-MCTS, a belief-state tree, in which each node is a belief-state, is constructed and search proceeds in accordance with beliefs. Then, Opponent Guessing and Opponent Predicting are proposed to illuminate the learning mechanism of beliefs with heuristic information. The beliefs are learned by heuristic information during search by specific methods, and we propose Opponent Guessing and Opponent Predicting to illuminate the learning mechanism. Besides, some possible improvements of the framework are investigated, such as incremental updating and all moves as first (AMAF) heuristic. Technical details are demonstrated about applying BS-MCTS to Phantom Go, especially on inference strategy. We examine the playing strength of the BS-MCTS and AMAF-BS-MCTS in Phantom Go by varying search parameters, also testify the proposed improvements.

Automatically acquiring domain knowledge from professional game records, a kind of pattern learning, is an attractive and challenging issue in computer Go. This paper proposes a supervised learning method, by introducing a new generalized Bradley-Terry model, named Only-One-Victor, to learn patterns from game records. Basically, our algorithm applies the same idea with Elo rating algorithm, which considers each move in game records as a group of move patterns, and the selected move as the winner of a kind of competition among all groups on current board. However, being different from the generalized Bradley-Terry model for group competition used in Elo rating algorithm, Only-One-Victor model in our work simulates the process of making selection from a set of possible candidates by considering such process as a group of independent pairwise comparisons. We use a graph theory model to prove the correctness of Only-One-Victor model. In addition, we also apply the Minorization-Maximization (MM) to solve the optimization task. Therefore, our algorithm still enjoys many computational advantages of Elo rating algorithm, such as the scalability with high dimensional feature space. With the training set containing 115,832 moves and the same feature setting, the results of our experiments show that Only-One-Victor outperforms Elo rating, a well-known best supervised pattern learning method.