Reusing previous task solutions in multiagent reinforcement learning

• Grant number: 18/00344-5
• Support type: Scholarships abroad - Research Internship - Doctorate
• Effective date (Start): April 01, 2018
• Effective date (End): March 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal researcher: Anna Helena Reali Costa
• Grantee: Felipe Leno da Silva
• Supervisor abroad: Peter Stone
• Home Institution: Escola Politécnica (EP). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Research place: University of Texas at Austin (UT), United States
• Associated to the scholarship: 15/16310-4 - Transfer Learning in Reinforcement Learning Multi-Agent Systems, BP.DR

Abstract

Reinforcement Learning (RL) is a popular solution to train autonomous agents to solve tasks in environments with unknown dynamics. Despite the recent successes of RL in solving increasingly challenging tasks, it suffers from an excessive requirement of samples of interactions with the environment in order to learn a good solution. When applied to Multiagent Systems (MAS), RL allows to train agents to solve tasks while coordinating with other agents. However, the sample complexity of MAS solutions is even higher, and hence Multiagent RL is dependent on additional techniques to be applicable to complex tasks. We here rely on Transfer Learning methods to accelerate Multiagent RL through knowledge reuse. In the main Ph.D. project we proposed a framework in which agents reuse knowledge both from previously learned tasks and advice from other agents. The main purpose of this exchange project is to develop methods to reuse previous task solutions in order to later integrate them in theafore mentioned framework. Recently, the use of Curriculum Learning in Single-Agent RL achieved interesting benefits for the learning process, accelerating the learning process of complex tasks. We here intend to contribute Curriculum Learning algorithms specialized to Multiagent RL. This is a resubmission of process 2017/13729-0, approved by FAPESP but not realized because of the long time required for issuing a US visa. (AU)

Scientific publications

(4)

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

DA SILVA, FELIPE LENO; REALI COSTA, ANNA HELENA. A Survey on Transfer Learning for Multiagent Reinforcement Learning Systems. JOURNAL OF ARTIFICIAL INTELLIGENCE RESEARCH, v. 64, p. 645-703, 2019. (16/21047-3, 18/00344-5, 15/16310-4)

DA SILVA, FELIPE LENO; NISHIDA, CYNTIA E. H.; ROIJERS, DIEDERIK M.; COSTA, ANNA H. REALI. Coordination of Electric Vehicle Charging Through Multiagent Reinforcement Learning. IEEE TRANSACTIONS ON SMART GRID, v. 11, n. 3, p. 2347-2356, MAY 2020. (15/16310-4, 18/00344-5)

GLATT, RUBEN; DA SILVA, FELIPE LENO; DA COSTA BIANCHI, REINALDO AUGUSTO; REALI COSTA, ANNA HELENA. DECAF: Deep Case-based Policy Inference for knowledge transfer in Reinforcement Learning. EXPERT SYSTEMS WITH APPLICATIONS, v. 156, OCT 15 2020. (16/21047-3, 15/16310-4, 18/00344-5, 16/18792-9)

DA SILVA, FELIPE LENO; WARNELL, GARRETT; COSTA, ANNA HELENA REALI; STONE, PETER. Agents teaching agents: a survey on inter-agent transfer learning. AUTONOMOUS AGENTS AND MULTI-AGENT SYSTEMS, v. 34, n. 1, APR 2020. (15/16310-4, 18/00344-5)