LiU > IDA > Real-Time Systems Lab


[26 May 2014] A bachelor student at RTSLAB was awarded the best thesis award from IDA - Simon Andersson. more ...

[31 May 2012] A masters student at RTSLAB was awarded the best thesis award from IDA - Ulf Magnusson. more ...

[27 February 2008] A masters student at RTSLAB was awarded the best thesis award from IDA - Johan Sigholm. more ...

[03 March 2004] A masters student at RTSLAB was awarded the best thesis award from IDA - Tobias Chyssler. more ...

[01 Jul 2003] For second year in a row a masters student at RTSLAB was awarded the best thesis award from SNART - Mehdi Amirijoo. more ...

Master Thesis - Past Projects - Abstract

Rule System with Adaptive Trigger Conditions

ID: LiTH-IDA-Ex-02/54

Controlling autonomous aircraft is an area within artificial intelligence that is studied at Saab AB in Linköping. The behaviour of a simulated autonomous aircraft can be decided by a system of rules, which is normally formulated as a set of conditional statements that respond to certain values in sensor signals from the environment. This thesis designs and evaluates a method for introducing learning abilities to the rule system that exercises control over a number of aircraft in a flight simulation. The rules controlling the aircraft are composed of three parts -N a triggering condition, an action and a consequence condition. When the triggering condition becomes true the action may be executed. The consequence condition specifies what an execution of the rule must fulfill to be considered successful. The learning approach described here assumes a system that receives an initial set of rules from a programmer, and improves its behaviour by tuning the trigger conditions of those rules. Every decision that the system makes is evaluated by comparing the actual result from executing the rule with the intended result, which is stated by the consequence condition. The resulting rate of success is stored, so that the accuracy of the system's response to succeeding situations can be improved by using the statistical data resulting from earlier decisions. A set of test cases, focusing on formation flight, evaluates the learning abilities of the system. For each of these cases, the system is given an initial set of rules in which many rules are often applicable to the same situation. Rule systems, like expert systems, practically let their rules execute in parallel, unless conflicts arise. By trying each of those rules that are competing for being used in a specific situation, this system can learn which rule gives the best performance for the current conditions. Exploiting this feature makes the task of constructing rules less time consuming, since many alternative rules can be initially applied to situations where it is not obvious to the rule-programmer which one will perform the best. The thesis concludes that the implemented rule system meets the basic requirements for adaptivity and expressiveness in the area of basic formation flight. Solutions generated by the current system are not optimal, but the report suggests some methods for significantly improving its performance.


Author(s): Knut Nordin

Contact: Nancy Reed

Click here to return.
Last modified September 2012. If you have questions or suggestions for the webpages, contact the webmaster