• Sorted by Date • Classified by Publication Type • Classified by Topic • Grouped by Student (current) • Grouped by Former Students •
Elad Mintzer. Agent Behavior Modeling by Learning from Action Logs. Master's Thesis, Bar Ilan University,2022.
The development of AI agents is a challenging task that is also time-consuming. ML techniques try to tackle this problem, allowing by interacting with the agent to its environment. However in our test case, there is no ready access to the environment, and the only data available is recorded data - logs. The data may not contain information about how the agent perceived its environment before taking action. Moreover, the data will often be missing information on the agent's internal processes because the logs record only globally observable information.This thesis tackles two challenges when modeling the behavior from logs recorded in continuous environments with continuous actions. In the first part of the thesis, we focus on the semi-automated learning of continuous action parameters. The method relies on guidance from a human domain expert but uses machine learning algorithms to carry out the actual learning. In the second part of the thesis, we focus on mining sequences of complex actions that appear in the logs. We build on earlier work in hierarchical sequence mining to introduce a novel method for mining action sequences where actions are complex and have discretized parameters.We hope the combination of the techniques from the two parts will lead towards the capacity for buildingfuller agent behavior models from logs of actions and environment settings.
@mastersthesis{mintzer-msc,
author = {Elad Mintzer},
title = { Agent Behavior Modeling by Learning from Action Logs},
school = {{B}ar {I}lan {U}niversity},
year = {2022},
OPTkey = {},
OPTtype = {},
OPTaddress = {},
OPTmonth = {},
OPTnote = {Available at \url{http://www.cs.biu.ac.il/~galk/Publications/b2hd-mintzer-msc.html}},
OPTannote = {},
wwwnote = {},
abstract = {
The development of AI agents is a challenging task that is also time-consuming. ML techniques try to tackle this problem, allowing by interacting with the agent to its environment. However in our test case, there is no ready access to the environment, and the only data available is recorded data - logs.
The data may not contain information about how the agent perceived its environment before taking action. Moreover, the data will often be missing information on the agent's internal processes because the logs record only globally observable information.
This thesis tackles two challenges when modeling the behavior from logs recorded in continuous environments with continuous actions. In the first part of the thesis, we focus on the semi-automated learning of continuous action parameters. The method relies on guidance from a human domain expert but uses machine learning algorithms to carry out the actual learning. In the second part of the thesis, we focus on mining sequences of complex actions that appear in the logs. We build on earlier work in hierarchical sequence mining to introduce a novel method for
mining action sequences where actions are complex and have discretized parameters.
We hope the combination of the techniques from the two parts will lead towards the capacity for building
fuller agent behavior models from logs of actions and environment settings.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Fri Aug 30, 2024 17:29:51