Multi-objective reinforcement learning methods for action selection : dealing with multiple objectives and non-stationarity

• Main Author: Anquise, Candy Alexandra Huanca
• Publication Date: 2021
• Format: Master thesis
• Language: eng
• Source: Biblioteca Digital de Teses e Dissertações da UFRGS
• Download full: http://hdl.handle.net/10183/231836
• topic: Sistemas multiagentes; Aprendizagem; Multi-objective; Decision-making; Multi-objective route choice; Reinforcement learning

Multi-objective decision-making entails planning based on a model to find the best policy to solve such problems. If this model is unknown, learning through interaction provides the means to behave in the environment. Multi-objective decision-making in a multi-agent system poses many unsolved challenges. Among them, multiple objectives and non-stationarity, caused by simultaneous learners, have been addressed separately so far. In this work, algorithms that address these issues by taking strengths from different methods are proposed and applied to a route choice scenario formulated as a multi-armed bandit problem. Therefore, the focus is on action selection. In the route choice problem, drivers must select a route while aiming to minimize both their travel time and toll. The proposed algorithms take and combine important aspects from works that tackle only one issue: non-stationarity or multiple objectives, making possible to handle these problems together. The methods used from these works are a set of Upper-Confidence Bound (UCB) algorithms and the Pareto Q-learning (PQL) algorithm. The UCB-based algorithms are Pareto UCB1 (PUCB1), the discounted UCB (DUCB) and sliding window UCB (SWUCB). PUCB1 deals with multiple objectives, while DUCB and SWUCB address non-stationarity in different ways. PUCB1 was extended to include characteristics from DUCB and SWUCB. In the case of PQL, as it is a state-based method that focuses on more than one objective, a modification was made to tackle a problem focused on action selection. Results obtained from a comparison in a route choice scenario show that the proposed algorithms deal with non-stationarity and multiple objectives, while using a discount factor is the best approach. Advantages, limitations and differences of these algorithms are discussed.