Home Research BCCN I Project Details Decision making and Planning
Personal tools

Decision making and Planning

Computational Neuroscience

The goal of this project is to design algorithms, which would allow an agent to discover the rules by which its environment is governed and to apply these rules in a goal-directed way. Clear emphasis lies on the question: How can an agent understand its world? Hence, how can it actually discover rules that couple perception with action in a reproducible way, thus, with a predictable outcome?


The goal is, thus, to extend the knowledge base of the agent by goal-directed exploration as well as supervised learning. The process, which we are investigating, relies on the idea of “surprise” as the triggering event for an extension of the knowledge base. As long as the agent finds that its actions lead to predictable outcome there is no surprise. A reproducibly unpredictable outcome of an action should trigger a process of “trying to resolve this surprise” (by learning), where the resolution of the surprise will enter a new entity into the agent knowledge base.


Two scenarios are being investigated:

  1. A real robot scenario in a block world where the robot’s goal is to discover which rules apply to the removal of different obstacle-combinations along its path. This scenario gets its complexity from the contingencies arising from the actually performed actions of the machine.

  2. SOKOBAN (look it up on the web if you do not know this game). This is a pure simulation scenario where the agent ought to discover the rules of this game. This scenario gets its complexity from the quite complex meta-rules which govern this game and should be discovered.


In the process of discovering rules the agent will gradually build a structure similar to a decision-tree, were these two projects are meant to focus of the process of building the tree and much less on the process of actually using the tree for planning.


Computational Neuroscience

Main cooperation partner:



Belongs to Group(s):