[1M4-GS-5x] Agents

4:40 PM - 4:55 PM JST(7:40 AM - 7:55 AM UTC)

[1M4-GS-5x-05] Adaptive Task Scheduling under Non-Stationary Resource Constraints: Exponential Fatigue Cost Functions and MaxEnt IRL for Dynamic Preference Estimation

〇Masakazu Yamaguchi¹ (1. Hibi Sohatsu)

[[online]]

Keywords:

Human-Agent Interaction,Adaptive Scheduling,Inverse Reinforcement Learning,Exponential Cost Function,Resource Constraints

This paper addresses action planning optimization in Human-Agent Interaction (HAI) systems where user internal resources, such as physical stamina, fluctuate non-stationarily. Conventional static scheduling fails to adapt to these dynamic constraints, leading to resource exhaustion. We propose an adaptive task scheduling architecture integrating two mathematical models. First, we formulate a "Dynamic Fatigue Cost Model" where execution costs increase exponentially with resource depletion. This acts as a soft control barrier function, preventing critical exhaustion. Second, to capture context-dependent values, we employ Maximum Entropy Inverse Reinforcement Learning (MaxEnt IRL). By assuming a Boltzmann distribution for action selection, the system dynamically estimates the reward function weights from behavioral logs, learning latent preferences. Discrete event simulations demonstrate that the agent autonomously detects cost spikes during resource downturns and pivots to low-load alternative tasks (Minimum Viable Products). This switching behavior smooths consumption curves, maximizing the long-term objective function. The framework offers a robust engineering basis for non-medical behavior change support, independent of clinical logic.