Applications of Planning and Learning

Classifier Learning for Medical and Technical Diagnosis

Kooperation mit Fraunhofer IIS, Medizinische Bildverarbeitung
Tom Hecker, Jörg Mennicke
Tom Hecker, Jörg Mennicke: Diagnosing Cancerous Abnormalities with Decision Tree Learning, [pdf] Publications
 Prototypische Realisierung eines Systems zum Erwerb und zur Anwendung von Expertenregeln [doc] thesis topic

Cognitive Assistance for People with Mental Deficiencies

Depending on the degree of mental impairment, everyday life of older people might be assisted in different ways: For simple forms of forgetfulness, people might use a voice-recording device keeping track of appointments, to-do lists, and shopping lists which can be either retrieved at demand or set to automatical reminding at specific times. If people gradually lose their ability to keep track of current goals, a monitoring and reminding system might give additional support to master the demands of daily life. Such a system should meet the following requirements: (1) Interleaving current plans with fixed routines in a convenient way (e.g. reminding a person that he/she needs to take some pills should be worked in between high-level goals of a plan and not in the middle of some sequence of connected actions). (2) Allowing people as much autonomy as possible, that is, being as less intrusive as possible (e.g., recognizing and allowing for goal switches and only prescribing actions if the person seems lost at what to do). (3) Being easy to use, especially not relying on explicit input from the user but instead {\em inferring} a current goal from interpreted sensor recordings of behavioral data. (4) Adaptivity to different grades of impairments (e.g. just reminding a person of top-level goals as eating lunch vs. guiding a person step-by-step through some sequence of actions).

Goal of the project is to develop an adaptible plan monitoring and guidance system fulfilling the requirements described above. Obviously, the system must be restricted with respect to such classes of planning tasks for which (1) sensor recordings are available and (2) behavioral intentions can be inferred from the sensor data. Nevertheless, we aim at developing a general framework which can be extendend to cover a larger set of ADLs if progress on the levels of sensor recordings and intention inference is made.

Collaborative distributed problem solving in unknown environments

Forschungsförderung 2007: Verteiltes Problemlösen in Notfallsituationen. Projekt-Nr. 070121-71 Grant

Learning Geographic Maps from Navigation

We developped a simulation system where an agent navigates a road system of arbitrary structure with many/sparse/no landmarks to find some specified goal location. During navigation the agent builds an internal representation of his navigation experience in form of a route graph. We are especially interested in algorithms for integrating route graphs from different egocentric perspectives into one allocentric map and in investigating efficiency trade-offs between navigation from scratch and building and using partial internal maps in dependence of the kind of road structures (rich/sparse graphs, many/few landmarks) and sensoric abilities of the agent (full metric knowledge, distorted observation of distances/angles, only knowledge about topologiy). Description
current: past:
  • Christopher Lörken
  • Christopher Lörken (bachelor student, cognitive science, Osnabrück): Learning Symbolic Maps from Robot Navigation (July 2004) [pdf, 70 pages - 510 KB] [Abstract]

Distributed/Local Planning for Transport Based on Ubiquituous Computing Techniques

Local algorithms for ubiquituous planning are developped and evaluated in the context of transportation logistics for intelligent streams of commodity. Clusters of goods, containers, warehouses, and transport vehicles, each equipped with ubiquituous computing technology, are considered as autonomous agents. In addition to these artificial agents, human agents (e.g. responsible for the logistics of a company) should be allowed to influence the constructed plans. Based on a multi-agent architecture (such as the belief-desire-intention model), the artificial agents shall generate local admissible plans where constraints given by human planners can be taken into consideration. The integration of local plans to (globally admissible) overall plans shall be performed bottom-up over several levels of hierarchy -- from clusters of goods, over transport vehicles to larger entities in the supply chain. Plan integration shall be performed by negotiation and communication. Furthermore, psychological studies on distributed problem solving of humans shall be performed, to provide suggestions for negotiation strategies which can be integrated as heuristics in the planning algorithms. In addition, planning algorithms based on human-like strategies should result in more easy to use assistant systems. Description