HTN Method Selection Guided by Multi-Objective Reinforcement Learning and Dynamic Preferences
9 Nov 2026
In the proceedings of ERAMIA-RS 2026 II Escola Regional de Aprendizado de Máquina e Inteligência Artificial da Região Sul (SBC) Proceedings
Anderson G. A. C. Filho (the name I sign papers with), Anderson R. Tavares and Lucas N. Alegre, Instituto de Informática, UFRGS. Four pages, in English. The version in the proceedings is here as a PDF, and so is the poster from ECOA 2026.
Abstract
This work proposes a hybrid framework in which Hierarchical Task Network (HTN) method selection is modeled as a temporally extended hierarchical decision, enabling the selected method to receive credit according to multiple objectives, each defined by a different reward function. This work aims to investigate whether the integration of symbolic planning and multi-objective reinforcement learning can provide adaptive agents while preserving interpretability, designer control, and explicit semantic constraints. The planned evaluation compares the proposed approach against traditional HTN method-selection strategies and scalar reinforcement learning baselines, considering adaptation to different preferences, generalization, semantic validity, and inference cost.
A small example
Take one decision with two objectives, progress and survival, and two applicable methods whose estimated returns are [0.8, 0.2] for m1 and [0.5, 0.5] for m2. With the weights [0.8, 0.2] the agent picks m1. Move the weights to [0.2, 0.8] and it picks m2. The HTN domain is the same in both cases. The numbers are made up to show the mechanism; no experiment has produced them yet.
Where it stands
This is the first version of the work. It describes the method and how it will be evaluated, and stops there. The experiments come next, in a controlled game-like environment, against first-applicable and heuristic HTN selection and two scalar RL baselines. With results in hand, the full paper goes to a conference in 2027. The research page has the longer explanation and the poster.
Cite
@inproceedings{filho2026htn,
author = {Anderson G. A. C. Filho and Anderson R. Tavares and Lucas N. Alegre},
title = {{HTN} Method Selection Guided by Multi-Objective Reinforcement Learning and Dynamic Preferences},
booktitle = {Anais da II Escola Regional de Aprendizado de M{\'a}quina e Intelig{\^e}ncia Artificial da Regi{\~a}o Sul (ERAMIA-RS)},
address = {Porto Alegre, RS},
publisher = {SBC},
year = {2026},
pages = {13--16},
doi = {10.5753/eramiars.2026.35023},
}