Discussing the article: "Neural Networks in Trading: Hierarchical Vector Transformer (HiVT)"

[MetaQuotes]

Check out the new article: Neural Networks in Trading: Hierarchical Vector Transformer (HiVT).

We invite you to get acquainted with the Hierarchical Vector Transformer (HiVT) method, which was developed for fast and accurate forecasting of multimodal time series.

The challenges in autonomous driving significantly overlap with those faced by traders. Navigating dynamic environments with safe maneuvers is a critical task for autonomous vehicles. To achieve this, these vehicles must comprehend their surroundings and predict future events on the road. However, accurately forecasting the maneuvers of nearby road users, such as cars, bicycles, and pedestrians, is a complex problem, particularly when their goals or intentions remain unknown. In multi-agent traffic scenarios, an agent's behavior is shaped by intricate interactions with other agents, further complicated by map-dependent traffic rules. Understanding the diverse behaviors of multiple agents in a scene is, therefore, extremely challenging.

Recent research uses a vectorized approach for more compact scene representation by extracting sets of vectors or points from trajectories and map elements. However, existing vectorized methods struggle with real-time motion prediction in fast-changing traffic conditions. Because such methods are usually sensitive to coordinate system shifts. To mitigate this issue, scenes are normalized to center on the target agent and align with its direction of movement. This approach becomes problematic when predicting the motion of a large number of agents, as the high computational costs of repeated scene normalization and feature recomputation for each target agent become a bottleneck. Additionally, existing methods model the relationships between all elements across spatial and temporal dimensions to capture detailed interactions between vectorized elements. This inevitably leads to excessive computational overhead as the number of elements increases. Since accurate real-time prediction is critical for autonomous driving safety, many researchers are looking to take this process to the next level by developing a new framework that enables faster and more precise multi-agent motion forecasting.

One such approach was presented in the paper "HiVT: Hierarchical Vector Transformer for Multi-Agent Motion Prediction". This method leverages symmetries and a hierarchical structure for multi-agent motion prediction. The authors of HiVT decompose the motion prediction task into multiple stages and hierarchically model interactions between elements using a Transformer-based architecture.

Author: Dmitriy Gizlyk