Extracting Knowledge from Reactive Robot Behaviour

Extracting Knowledge from Reactive Robot Behaviour. Ledezma, A., Berlanga, A., & Aler, R. In Proceedings of the Agents-01 Workshop on Learning Agents, pages 7-12, Montreal, Canada, May, 2001.
abstract bibtex

In previous work, we have presented an approach that allows to acquire a declarative representation of the behaviour of an agent, by observing what output it produces from its inputs. So is, we acquire a model of other agents. In that case, only domains coming from the UCI database were tried. In this paper, we test our approach to model the behavior of an actual agent: a simulated robot whose controller (a feed-forward neural network) was learned by using new coevolutionary techniques. In our previous work, we used C4.5 to model other agents. However, C4.5 cannot handle continuous classes, hence they have to be discretized. Besides reporting the results of applying C4.5+discretization to the robot domain, we extend our approach by using a machine learning technique able to use continuous outputs (M5). Finally, we compare the learned models to the neural net controller by allowing those models to control the robot directly with the models obtained from C4.5 and M5 in actual simulations.

@INPROCEEDINGS{extracting-agent01,
  author = {Agapito Ledezma and Antonio Berlanga and Ricardo Aler},
  title = {Extracting Knowledge from Reactive Robot Behaviour},
  booktitle = {Proceedings of the Agents-01 Workshop on Learning Agents},
  year = {2001},
  pages = {7-12},
  address = {Montreal, Canada},
  month = {May},
  abstract = {In previous work, we have presented an approach that allows to acquire
	
	a declarative representation of the behaviour of an agent, by observing
	
	what output it produces from its inputs. So is, we acquire a model
	
	of other agents. In that case, only domains coming from the UCI
	
	database were tried. In this paper, we test our approach to model
	the
	
	behavior of an actual agent: a simulated robot whose controller (a
	
	feed-forward neural network) was learned by using new coevolutionary
	
	techniques. In our previous work, we used C4.5 to model other
	
	agents. However, C4.5 cannot handle continuous classes, hence they
	
	have to be discretized. Besides reporting the results of applying
	
	C4.5+discretization to the robot domain, we extend our approach by
	
	using a machine learning technique able to use continuous outputs
	
	(M5). Finally, we compare the learned models to the neural net
	
	controller by allowing those models to control the robot directly
	with
	
	the models obtained from C4.5 and M5 in actual simulations.},
  bib2html_pubtype = {Workshop},
  bib2html_rescat = {Agent Modeling},
  days = {29},
  key = {modelado},
  owner = {ledezma},
  timestamp = {2011.11.21}
}

Downloads: 0

{"_id":{"_str":"5342a9600e946d920a002f66"},"__v":11,"authorIDs":["54575aa32abc8e9f370002ec","54576e722abc8e9f370003c4","5463c0594f4818cf4f0003de","8mdzoMZ27yguKiBHc"],"author_short":["Ledezma, A.","Berlanga, A.","Aler, R."],"bibbaseid":"ledezma-berlanga-aler-extractingknowledgefromreactiverobotbehaviour-2001","bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"firstnames":["Agapito"],"propositions":[],"lastnames":["Ledezma"],"suffixes":[]},{"firstnames":["Antonio"],"propositions":[],"lastnames":["Berlanga"],"suffixes":[]},{"firstnames":["Ricardo"],"propositions":[],"lastnames":["Aler"],"suffixes":[]}],"title":"Extracting Knowledge from Reactive Robot Behaviour","booktitle":"Proceedings of the Agents-01 Workshop on Learning Agents","year":"2001","pages":"7-12","address":"Montreal, Canada","month":"May","abstract":"In previous work, we have presented an approach that allows to acquire a declarative representation of the behaviour of an agent, by observing what output it produces from its inputs. So is, we acquire a model of other agents. In that case, only domains coming from the UCI database were tried. In this paper, we test our approach to model the behavior of an actual agent: a simulated robot whose controller (a feed-forward neural network) was learned by using new coevolutionary techniques. In our previous work, we used C4.5 to model other agents. However, C4.5 cannot handle continuous classes, hence they have to be discretized. Besides reporting the results of applying C4.5+discretization to the robot domain, we extend our approach by using a machine learning technique able to use continuous outputs (M5). Finally, we compare the learned models to the neural net controller by allowing those models to control the robot directly with the models obtained from C4.5 and M5 in actual simulations.","bib2html_pubtype":"Workshop","bib2html_rescat":"Agent Modeling","days":"29","key":"extracting-agent01","owner":"ledezma","timestamp":"2011.11.21","bibtex":"@INPROCEEDINGS{extracting-agent01,\r\n author = {Agapito Ledezma and Antonio Berlanga and Ricardo Aler},\r\n title = {Extracting Knowledge from Reactive Robot Behaviour},\r\n booktitle = {Proceedings of the Agents-01 Workshop on Learning Agents},\r\n year = {2001},\r\n pages = {7-12},\r\n address = {Montreal, Canada},\r\n month = {May},\r\n abstract = {In previous work, we have presented an approach that allows to acquire\r\n\t\r\n\ta declarative representation of the behaviour of an agent, by observing\r\n\t\r\n\twhat output it produces from its inputs. So is, we acquire a model\r\n\t\r\n\tof other agents. In that case, only domains coming from the UCI\r\n\t\r\n\tdatabase were tried. In this paper, we test our approach to model\r\n\tthe\r\n\t\r\n\tbehavior of an actual agent: a simulated robot whose controller (a\r\n\t\r\n\tfeed-forward neural network) was learned by using new coevolutionary\r\n\t\r\n\ttechniques. In our previous work, we used C4.5 to model other\r\n\t\r\n\tagents. However, C4.5 cannot handle continuous classes, hence they\r\n\t\r\n\thave to be discretized. Besides reporting the results of applying\r\n\t\r\n\tC4.5+discretization to the robot domain, we extend our approach by\r\n\t\r\n\tusing a machine learning technique able to use continuous outputs\r\n\t\r\n\t(M5). Finally, we compare the learned models to the neural net\r\n\t\r\n\tcontroller by allowing those models to control the robot directly\r\n\twith\r\n\t\r\n\tthe models obtained from C4.5 and M5 in actual simulations.},\r\n bib2html_pubtype = {Workshop},\r\n bib2html_rescat = {Agent Modeling},\r\n days = {29},\r\n key = {modelado},\r\n owner = {ledezma},\r\n timestamp = {2011.11.21}\r\n}\r\n\r\n","author_short":["Ledezma, A.","Berlanga, A.","Aler, R."],"id":"extracting-agent01","bibbaseid":"ledezma-berlanga-aler-extractingknowledgefromreactiverobotbehaviour-2001","role":"author","urls":{},"downloads":0,"html":"","metadata":{"authorlinks":{"ledezma, a":"http://www.caos.inf.uc3m.es/ledezma/ca/publicaciones/"}}},"bibtype":"inproceedings","biburl":"http://www.caos.inf.uc3m.es/bibs/pubCAOS.bib","downloads":0,"keywords":[],"search_terms":["extracting","knowledge","reactive","robot","behaviour","ledezma","berlanga","aler"],"title":"Extracting Knowledge from Reactive Robot Behaviour","year":2001,"dataSources":["Z4T4fewybNpYCWveD"]}