Hierarchical clustering is a fundamental machine-learning technique for grouping data points into dendrograms. However, existing hierarchical clustering methods encounter two primary challenges: 1) Most methods specify dendrograms without a global objective. 2) Graph-based methods often neglect the significance of graph structure, optimizing objectives on complete or static predefined graphs. In this work, we propose Hyperbolic Continuous Structural Entropy neural networks, namely HypCSE, for structure-enhanced continuous hierarchical clustering. Our key idea is to map data points in the hyperbolic space and minimize the relaxed continuous structural entropy (SE) on structure-enhanced graphs. Specifically, we encode graph vertices in hyperbolic space using hyperbolic graph neural networks and minimize approximate SE defined on graph embeddings. To make the SE objective differentiable for optimization, we reformulate it into a function using the lowest common ancestor (LCA) on trees and then relax it into continuous SE (CSE) by the analogy of hyperbolic graph embeddings and partitioning trees. To ensure a graph structure that effectively captures the hierarchy of data points for CSE calculation, we employ a graph structure learning (GSL) strategy that updates the graph structure during training. Extensive experiments on seven datasets demonstrate the superior performance of HypCSE.

Vision-Language Transformers (VLTs) have shown great success recently, but are meanwhile accompanied by heavy computation costs, where a major reason can be attributed to the large number of visual and language tokens. Existing token pruning research for compressing VLTs mainly follows a single-modality-based scheme yet ignores the critical role of aligning different modalities for guiding the token pruning process, causing the important tokens for one modality to be falsely pruned in another modality branch. Meanwhile, existing VLT pruning works also lack the flexibility to dynamically compress each layer based on different input samples. To this end, we propose a novel framework named Multimodal Alignment-Guided Dynamic Token Pruning (MADTP) for accelerating various VLTs. Specifically, we first introduce a well-designed Multi-modality Alignment Guidance (MAG) module that can align features of the same semantic concept from different modalities, to ensure the pruned tokens are less important for all modalities. We further design a novel Dynamic Token Pruning (DTP) module, which can adaptively adjust the token compression ratio in each layer based on different input instances. Extensive experiments on various benchmarks demonstrate that MADTP significantly reduces the computational complexity of kinds of multimodal models while preserving competitive performance. Notably, when applied to the BLIP model in the NLVR2 dataset, MADTP can reduce the GFLOPs by 80% with less than 4% performance degradation. The code is available at https://github.com/double125IMADTP.

Multi-agent reinforcement learning (MARL) effectively improves the learning speed of agents in sparse reward tasks with the guide of subgoals. However, existing works sever the consistency of the learning objectives of the subgoal generation and subgoal reached stages, thereby significantly inhibiting the effectiveness of subgoal learning. To address this problem, we propose a novel Potential field Subgoal-based Multi-Agent reinforcement learning (PSMA) method, which introduces the potential field (PF) to unify the two-stage learning objectives. Specifically, we design a state-to-PF representation model that describes agents' states as potential fields, allowing easy measurement of the interaction effect for both allied and enemy agents. With the PF representation, a subgoal selector is designed to automatically generate multiple subgoals for each agent, drawn from the experience replay buffer that contains both individual and total PF values. Based on the determined subgoals, we define an intrinsic reward function to guide the agent to reach their respective subgoals while maximizing the joint action-value. Experimental results show that our method outperforms the state-of-the-art MARL method on both StarCraft II micro-management (SMAC) and Google Research Football (GRF) tasks with sparse reward settings.

In some complex multi-agent environments, the types of relationships between agents are diverse and their intensity changes during the policy learning process. Theoretically, some of these relationships can facilitate cooperative policy learning. However, acquiring these relationships is an intractable problem. To tackle the problem, we propose a diverse effective relationship exploration based multi-agent reinforcement learning (DERE) method. Specifically, a potential fields model is firstly designed to represent relationships between agents. Then to encourage the exploration of effective relationships, we define an information-theoretic objective function. Finally, an intrinsic reward function is designed to optimize the information-theoretic objective, meanwhile, guide agents to learn more effective collaborative policies. Experimental results show that our method outperforms state-of-the-art methods on both super hard StarCraft II micromanagement tasks (SMAC) and Google Research Football (GRF).

Pommerman is a recently-proposed multi-agent benchmark, which is very challenging for Reinforcement Learning (RL). The main obstacles to adopt RL in Pommerman are the delayed action effects and sparse rewards. This paper presents novel approaches to mitigate these problems by introducing Artificial Potential Field (APF) in the two-dimensional Pommerman world. We propose a new framework to generate hybrid features from both APF computation and raw environment data. Meanwhile, a new reward shaping method through APF is developed to give the learning agent faster and more efficient policy iteration. The training results show that the learning speed and convergence reward are both improved on a 1v1 mode of the Pommerman game, compared to the conventional learning algorithms, A2C and ACKTR.