Esta página em
I am a lecturer (Professor Auxiliar) at the FEUP (Engineering Faculty of Porto University). I am also researcher at LIACC (Artificial Intelligence and Computer Science Laboratory) where I am a member of the Machine Learning group. LIACC is located in CIUP (University of Porto Computer Centre). I am also a member of NIA&R. (Artificial Intelligence and Robotics group) at FEUP. I am a member of APPIA (Portuguese Association for Artificial Intelligence).
The use of Machine Learning (ML) tools to reverse-engineer a human control skill is usually done in three stages. First, a human 'pilot' executes the control task several times. During task execution, the state variables of the system being controlled are recorded at regular time intervals. Each such record forms a state vector. In the second stage, the sampled state vectors are taken as examples by the ML tool and a controller is generated. In the third stage, this machine-constructed controller replaces the human 'pilot' in the control task.
This thesis evaluates propositional and first order Machine Learning algorithms on the problem of reverse-engineering an F-16 human pilot.
The experiments concentrate mainly on extracting a set of basic flight manoeuvres. The ML tools being compared are C4.5, as a decision tree representative, and IndLog (under development), as a representative of an Inductive Logic Programming (ILP) tool.
Some limitations of current decision tree and ILP systems imply that the tools much be developed to make them adequate for this kind of reverse engineering.
Basic flying manoeuvres are modeled by a set of operators, the definitions of which are to be machine-induced. To build up such operators, decision trees have to be extended into what is called Parameterised Decision Trees. A Parameterised Decision Tree is a decision tree associated with a set of parameters (goals in the control domain). These parameters are used in computing the values of some attributes.
The ILP system IndLog is being developed to cope with some features of the flying domain that make the existing ILP system difficult to apply to this domain. These features include the existence of a large number of examples, the need to search large search spaces, and the ability to handle noisy and numerical data.
The experiments are done using ACM public domain flight simulator version
2.4 (acm-2.4.tgz 169K), although there
is a more recent version (acm-4.7.tgz 1050K).