Interpreting Tactical Decision-Making in Transformer-Based Agents

Description

The development of autonomous agents capable of sophisticated strategic decision-making in complex environments is a central goal of artificial intelligence. This paper proposes a framework for discovering and interpreting strategies in a simulated grid-world battlefield environment. We leverage the AlphaZero algorithm, a powerful reinforcement learning method that combines Monte Carlo Tree Search (MCTS) with deep neural networks, to train agents. Crucially, the neural network component incorporates a Transformer architecture. The primary contribution of this work lies in the proposed methodology to utilize the self-attention mechanisms within the Transformer to gain insights into the agent’s decision-making process, specifically by visualizing which parts of the battlefield the network pays attention to when selecting an action. This approach aims not only to develop high-performing agents but also to enhance the interpretability of their learned strategies.

Period	25 Nov 2025
Event title	1st International Conference on Big Data Analytics and Applications
Event type	Conference
Conference number	1
Location	Innsbruck, AustriaShow on map
Degree of Recognition	International