Andrés Occhipinti Liberman: The aim of this PhD project is to build artificial agents (e.g. robots) that can learn autonomously, through observation and communication, how to act in a certain environment to solve a given task.
Specifically, our goal is to endow the agents with the ability to learn action models. Loosely speaking, an action model is a structured description of how a certain action works; what it does (the action’s effects), and when it can be executed (the action’s preconditions). For example, for a simple action such as picking up a box, the preconditions may state that the box must not have another box on top of it, and the effects may state that the box will be held by the robot applying the action. When an agent has well-defined action models, it can use this information to achieve its goals. For instance, a robot that knows how to pick up boxes can then use this knowledge to move a set of boxes around a warehouse and get them into a desired configuration.
Action models are normally hand-coded by engineers; before placing a box-carrying robot in a warehouse, a team of engineers would hand-code how the robot can perform actions such as `picking up a box’ or ‘pushing a box’. Hand-crafting action models in this way is a difficult and error-prone task. Artificial agents are typically intended to act in large, dynamic, and unpredictable environments. For such environments, it is often unfeasible to fully specify action models in advance. Action model learning tries to solve this problem by endowing the agents with the ability to generate action models on their own. The agents may initially not know what the actions do, but they can try them out. They can then observe what happens and, after a certain amount of experimentation, they can try to infer how the actions work in general. This project is focused on designing algorithms to support this kind of learning.
Most existing methods for action model learning focus on single agents learning how to act in isolation from others. However, many real-world problems (security surveillance, search-and-rescue operations, etc.) are best solved by multiple agents acting in a shared environment. Such groups of agents are called multi-agent systems. Learning action models in settings with multiple interacting agents is qualitatively different from learning in a solitary way. This is because the learning activities of each individual agent may be significantly influenced (delayed, accelerated, or made possible at all) by other agents. As a result, the existing single-agent learning methods do not generally work when applied to multi-agent settings. Our main goal is to bridge this gap by designing algorithms for learning in multi-agent settings.
PhD project 2017 -
DTU supervisor: Thomas Bolander, section for Algorithms, Logic and Graphs at DTU Compute.