B10 – The Development of Natural Induction Methodology for Discovering Patterns Supporting Autonomous Transportation (new)

Motivation

This subproject will develop a methodology for automated discovery of traffic patterns from historic data characterizing the traffic in a given geographical region. The methodology will be based on natural induction approach that seeks computer knowledge in the forms natural to people, such as natural language-type descriptions and graphical representations. It will be implemented by extending the attributional rule learning system AQ21, and tested on the data provided by the CRC.


Objectives Phase 2 (2008-2011)

Goals of this sub project, which is conducted in collaboration with the CRC 637 comprise the development of a methodology for the automated discovery of traffic patterns from historic data characterizing the traffic in a given geographical region, the design and implementation of appropriate software tools and the systematic evaluation of the methodology by means of multiagent-based simulation experiments in the PlaSMA system.

In logistics, control knowledge learned by an agent of course needs to be accurate. In addition, it also needs to be understandable and potentially modifiable by a human process expert. The rationale for the latter requirements, which are rarely satisfied by current learning systems, is that informed decisions which are rendered in autonomous control need to be comprehensible for human controllers. The representation of knowledge and models in a human-oriented form is thereby essential for the ability of software agents to justify their decision-making. Also, process-relevant knowledge may be validated and potentially complemented by a process expert.

Approach Phase 2 (2008-2011)

To satisfy the aforementioned objectives to generate and represent knowledge in human-oriented forms, a particularly attractive learning approach is natural induction. It strives to discover patterns in presented data that are encoded in forms corresponding to simple natural-language statements.

The AQ21 rule learning system was utilized for discovery of traffic patterns in data collected by truck management agents over extended periods of time within a simulation environment. Conceived as a general-purpose tool to attributional classifiers from multi-class data, AQ21 needed to integrated with the PlaSMA simulation platform. The learning system also needed to be extended with regard to the handling of incomplete and ambiguous data. These steps then allowed for a detailed assessment of AQ21 learning performance and convergence behavior in logistic use cases.

In a complimentary strand of research, the Learnable Evolution Model (LEM), an approach for non-darwinian evolutionary optimization also developed at the MLI laboratory, was utilized as a means to optimize individual truck transport schedules.

Results Phase 2 (2008-2011)

With the subproject B4, PlaSMA-based experiments were conducted where transport agents independently tried and optimized their respective pickup&delivery route in terms of driving time based on different route assessment models. While some agents took into account only weather information, others additionally employed AQ21 to learn prediction models for avoidance of traffic congestions. Statistical evaluation showed that weather-aware agents were 5.3 % faster than situation-agnostic naïve agents. Agents with traffic prediction were shown to be 6.3% faster. Additionally, the standard deviation of driving times could be reduced significantly (from 52 to 36 minutes for a 1000 km trip).

Additional experiments with an emphasis on learning time for prediction models based on local experience showed that only one month worth of local observations was required for near-optimal models. AQ learning was then combined successfully with Q-learning achieve an even better convergence behavior.

Finally, the Learnable Evolution Model has been employed for the optimization of individual transport schedules.