IEEE International Conference on Intelligent Systems, London, 2010.
High level conceptual thought seems to be at the basis of the impressive human cognitive ability, and AI researchers aim to replicate this ability in artificial agents. Classical top-down (Logic based) and bottom-up (Connectionist) approaches to the problem have had limited success to date. We review a small body of work that represents a different approach to AI. We call this work the Bottom Up Symbolic (BUS) approach and present a new BUS method to concept construction. While valid concepts have been constructed using previous methods under this approach, we show in this paper that the one-sided clustering methods generally used there may fail to uncover valid concepts even when they clearly exist. We show that by using a Co-clustering algorithm that searches for an optimal partitioning based on the Mutual Information between the category and consequent components of a concept, the concept formation outcome is improved. We test our approach on data from experiments using a real mobile robot operating in the real world, and show that our Co-clustering based approach leads to significant performance improvement compared to previous approaches.
The International Conference on Intelligent Agent Technology, Milan, 2009
High level conceptual thought seems to be at the basis of the impressive human cognitive ability. Classical top-down (Logic based) and bottom-up (Connectionist) approaches to the problem have had limited success to date. We identify a small body of work that represents a different approach to AI. We call this work the Bottom Up Symbolic (BUS) approach and present a new BUS method to concept construction. The main novelty of our work is that we apply statistical methods in the concept construction process. Our findings here suggest that such methods are necessary since a symbolic description of the true agent-environment interaction dynamics is often hidden among a background of non-representative descriptions, especially if data from unconstrained real-world experiments is considered. We consider such data (from a mobile robot randomly roaming an office environment) and show how our method can correctly grow a set of true concepts from the data.
The Encyclopedia of Data Warehousing and Mining (2nd Edition) , Ed. J.Wang, IGI Global, 2008
Clustering analysis is a tool used widely in the Data Mining community and beyond (Everitt et al. 2001). In essence, the method allows us the information in a large data set X by creating a very much smaller set C of representative points (called centroids) and a membership map relating each point in X to its representative in C. An obvious but special type of data set that one might want to cluster is a time series data set. Such data has a temporal ordering on its elements, in contrast to non-time series data sets. In this article we explore the area of time series clustering, focusing mainly on a surprising recent result showing that the traditional method for time series clustering is meaningless. We then survey the literature of recent papers and go on to argue how time series clustering can be made meaningful.
The Australasian Data Mining Conference, Gold Coast, Australia, 2007
Clustering time series data using the popular subsequence (STS) technique has been widely used in the data mining and wider communities. Recently the conclusion was made that it is meaningless, based on the findings that it produces (a) clustering outcomes for distinct time series that are not distinguishable from one another, and (b) cluster centroids that are smoothed. More recent work has since showed that (a) could be solved by introducing a lag in the subsequence vector construction process, however we show in this paper that such an approach does not solve (b). Motivating the terminology that a clustering method which overcomes (a) is meaningful, while one which overcomes (a) and (b) is useful, we propose an approach that produces useful time series clustering. The approach is based on restricting the clustering space to extend only over the region visited by the time series in the subsequence vector space. We test the approach on a set of 12 diverse real-world and synthetic data sets and find that (a) one can distinguish between the clusterings of these time series, and (b) that the centroids produced in each case retain the character of the underlying series from which they came.
Knowledge and Information Systems, Springer
Sequential time series clustering is a technique used to extract important features from time series data. The method can be shown to be the process of clustering in the Delay-Vector space formalism used in the Dynamical systems literature. Recently, the startling claim was made that sequential time series clustering is meaningless. This has important consequences for a significant amount of work in the literature, since such a claim invalidates these work's contribution. In this paper, we show that sequential time series clustering is not meaningless, and that the problem highlighted in these works stem from their use of the Euclidean distance metric as the distance measure in the delay vector space. As a solution, we consider quite a general class of time series, and propose a regime based on two types of similarity that can exist between delay vectors, which give rise naturally to an alternative distance measure to Euclidean distance in the delay vector space. We show that, using this alternative distance measure, sequential time series clustering can indeed be meaningful.
Int. Conf. Data Mining, Houston Tex. Nov. 2005
Recently, the startling claim was made that sequential time series clustering is meaningless. This has important consequences for a significant amount of work in the literature, since such a claim invalidates this work's contribution. In this paper, we show that sequential time series clustering is not meaningless, and that the problem highlighted in these works stem from their use of the Euclidean distance metric as the distance measure in the subsequence vector space. As a solution, we consider quite a general class of time series, and propose a regime based on two types of similarity that can exist between subsequence vectors, which give rise naturally to an alternative distance measure to Euclidean distance in the subsequence vector space. We show that, using this alternative distance measure, sequential time series clustering can indeed be meaningful.
Int. J. Robotics Research, December 2005
Programming by Demonstration (PbD) is a technique for programming robots that holds much promise in making robots more accessible to ordinary, non-technical users. However, a well known difficulty with the method is that a human will often demonstrate the task to be programmed inconsistently or even erroneously, leading to the inclusion of what is essentially noise in the demonstration. A number of techniques exist in the literature for filtering out this type of noise, however most focus on very low level control command details. In this paper we propose a new, complimentary direction of research. We take a ``task-level'' view of the demonstration, and note that noise can exist at this level also. We propose a framework, based on a Hybrid Dynamic System modelling approach, to select the most optimal, task-level execution strategies that were demonstrated. We apply our framework to a real household task of inserting the compressible spindle of a paper towel holder into its supports. We conduct experiments to show that significant improvements in robot performance of the task can be achieved by a PbD regime that includes our method.
Int. J. Robotics Research, vol. 22, no. 5, pp. 299-319, May 2003
© 2003 Massachusetts Institute of Technology
The difficulty associated with programming existing robots is one of the main impediments to them finding application in domestic environments like the home. A promising method for simplifying robot programming is Programming by Demonstration (PbD). Here, an end user can provide a demonstration of the task to be programmed, with a PbD ``interface'' interpreting the demonstration in order to determine low-level control details for the robot. A key aspect of the interpretation process is to make it robust to the noise typically included in a demonstration by the human. In this paper we present a method to help identify and eliminate any noise present in the demonstration. Our method involves two steps. The first step uses the demonstration to build-up a partial knowledge of the geometry present in the task. Statistical regression analysis is used on demonstrated trajectories to determine equations describing C-surfaces in Configuration Space. The second step in our method uses the geometric information obtained in the first step to determine if there are more optimal paths than those demonstrated for completing the task. If there are, our method proposes these as the appropriate control commands for the robot. We show the validity of our approach by presenting successful experiments on a realistic household-type task - changing rolls on a paper roll holder.
Unpublished PhD Thesis, July 2001
Finding a simple but powerful robot programming method for realistic
tasks has been one of the main aims of robotics researchers for over
two decades. A promising approach is robot Programming by
Demonstration (PbD). Here, a demonstration of the task is interpreted
by a PbD interface so that a set of control commands to achieve the
task are produced for the robot. PbD is a promising approach to robot
programming, however a well known weakness of the method is that human
demonstrations can be suboptimal. Research has identified that
demonstrations can contain inconsistencies, noise, or even incorrect
or unintended actions. In this thesis our focus is on identifying and
removing sub-optimality from the demonstration. Our aim is to ensure
that control commands formed for the robot from demonstrated actions
encode efficient and reliable execution of the task. The work we propose
is divided into three distinct areas. They are: (i) determining
task-specific, geometric properties of a task from demonstration, (ii)
deriving efficient, low-level, robot control-commands from
demonstration, and (iii) determining optimal task-level strategies
Research area (i) is important since knowledge of task geometry can help identify the presence of suboptimal actions in the demonstration. Our solution uses the concept of Configuration Space (C-space) as a means to represent task geometry. We apply statistical regression analysis to build-up a knowledge of task geometry in regions of the task that were visited in the demonstration. Experimental results showed the validity of the approach. The two key results were, first, that only a partial representation of C-space could be determined. That is, a representation of C-space was determined only for regions of the task visited in the demonstration. Second, the method could determine quite accurately the true geometric properties of the task, so long as it had sufficient information to do so. That is, C-space was derived accurately in regions visited often in the demonstration, and less accurately in less-visited regions. %-what is a region of the task??
Research area (ii) involves deriving low-level control commands for the robot from demonstration, and has been the main focus of research into removing demonstration sub-optimality in PbD to date. In this thesis we adopt the well known control regime of hybrid force-position control, and so the problem divides into two sub-areas: position control-command synthesis, and force control-command synthesis. For position control-command synthesis we present a novel method for path planning in a partially known C-space. The method has the advantages compared to other methods in the literature that, it can derive paths containing undemonstrated points, it is applicable to a task with any degree of freedom, and that it does not assume a set form of demonstration topology. A drawback of the approach is that, theoretically, in some circumstances a valid control command will not be derived. However, we show with experimental results for a realistic task how with appropriate tuning, the method will produce a valid set of position control commands. For force control-command synthesis, a well known characteristic of force commands recorded from a demonstration are that: they include friction, and that force sensors introduce into these forces a high frequence noise component. We use our knowledge of C-space to determine an appropriate direction for force control. We base the magnitude of force control commands on the forces commanded by the human in the demonstration, however we remove friction and noise using filtering and spline fitting techniques.
Research area (iii) has to our knowledge not been the subject of work in the PbD field to date. In this thesis, we identify that a set of demonstrations will contain many different task-level strategies. Some of these will result in more optimal robot performance of the task than others. We present a method for selecting a task-level strategy for the robot. It is based on the concept of modeling skill in a task as a Hybrid Dynamic System. The Hybrid Dynamic System representation allows a demonstration to be represented at a task level as a sequence of discrete events. Our method is to form an automaton from the event sequences encoded by each demonstration. We use a set of metrics to determine a cost for each event in the automaton. Finally, an optimal task level strategy is produced by conducting a least cost path search in the automaton. We show with experiments how the task-level strategies selected by our method do in fact result in more optimal robot performance of the task.