David Rajaratnam

A Novel Architecture for Situation Awareness Systems

Situation Awareness (SA) is the problem of comprehending elements of an environment within a volume of time and space. It is a crucial factor in decision-making in dynamic environments. Current SA systems support the collection, filtering and presentation of data from different sources very well, and typically also some form of low-level data fusion and analysis, e.g., recognizing patterns over time. However, a still open research challenge is to build systems that support higher-level information fusion, viz., to integrate domain specific knowledge and automatically draw conclusions that would otherwise remain hidden or would have to be drawn by a human operator. To address this challenge, we have developed a novel system architecture that emphasizes the role of formal logic and automated theorem provers in its main components. Additionally, it features controlled natural language for operator I/O. It offers three logical languages to adequately model different aspects of the domain. This allows to build SA systems in a more declarative way than is possible with current approaches. From an automated reasoning perspective, the main challenges lay in combining (existing) automated reasoning techniques, from low-level data fusion of time-stamped data to semantic analysis and alert generation that is based on linear temporal logic. The system has been implemented and interfaces with Google-Earth to visualize the dynamics of situations and system output. It has been successfully tested on realistic data, but in this paper we focus on the system architecture and in particular on the interplay of the different reasoning components.

Integrating ASP into ROS for Reasoning in Robots

Knowledge representation and reasoning capacities are vital to cognitive robotics because they provide higher level functionalities for reasoning about actions, environments, goals, perception, etc. Although Answer Set Programming (ASP) is well suited for modelling such functions, there was so far no seamless way to use ASP in a robotic setting. We address this shortcoming and show how a recently developed ASP system can be harnessed to provide appropriate reasoning capacities within a robotic system. To be more precise, we furnish a package integrating the new version of the ASP solver clingo with the popular open-source robotic middleware Robot Operating System (ROS). The resulting system, ROSoClingo, provides a generic way by which an ASP program can be used to control the behaviour of a robot and to respond to the results of the robot’s actions.

A systematic solution to the (De-)composition problem in general game playing

General game players can drastically reduce the cost of search if they are able to solve smaller subproblems individually and synthesise the resulting solutions. To provide a systematic solution to this (de-)composition problem, we start off with generalising the standard decomposition problem in planning by allowing the composition of individual solutions to be further constrained by domain-dependent requirements of the global planning problem. We solve this generalised problem based on a systematic analysis of composition operators for transition systems, and we demonstrate how this solution can be further generalised to general game playing.

Prime implicates for approximate reasoning

Techniques for improving the computational efficiency of inference have held a long fascination in computer science. Two popular methods include approximate logics and knowledge compilation. In this paper we apply the idea of approximate compilation to develop a notion of prime implicates for the family of classically sound, but incomplete, approximate logics S-3. These logics allow for differing levels of approximation by varying membership of a set of propositional atoms. We present a method for computing the prime S-3-implicates of a clausal knowledge base and empirical results on the behaviour of prime S-3-implicates over randomly generated 3-SAT problems. A very important property of S- 3-implicates and our algorithm for computing them is that decreasing the level of approximation can be achieved in an incremental manner without re-computing from scratch (Theorem 7).

Composability in Cognitive Hierarchies

This paper develops a theory of node composition in a formal framework for cognitive hierarchies. It builds on an existing model for the integration of symbolic and sub-symbolic representations in a robot architecture consisting of nodes in a hierarchy. A notion of behaviour equivalence between cognitive hierarchies is introduced and node composition operators that preserve this equivalence are defined. This work is significant in two respects. Firstly, it opens the way for a formal comparison between cognitive robotic systems. Secondly, composition, more precisely decomposition, has been shown to be important to many fields, and may therefore prove of practical benefit in the context of cognitive systems.

From approximate clausal reasoning to problem hardness

Approximate propositional logics provide a response to the intractability of classical inference for the modelling and construction of resource-bounded agents. They allow the degree of logical soundness (or completeness) to be balanced against the agents resource limitations. We develop a logical semantics, based on a restriction to Fingers logics of limited bivalence [5], and establish the adequacy of a clausal tableau based proof theory with respect to this semantics. This system is shown to characterise DPLL with restricted branching, providing a clear path for the adaptation of DPLL-based satisfiability solvers to approximate reasoning. Furthermore it provides insights into the traditional notion of problem hardness, as we show that the parameter set of these logics correspond to the strong backdoor for an unsatisfiable problem.