Evolving spatial and temporal lexicons across different cognitive architectures [thesis]

Scott Heath
Communication between mobile robots requires a transfer of symbols, where each symbol signifies a meaning. However, in typical applications, meaning has been ascribed to the symbols by the engineers that have programmed the robots. This thesis explores an alternative: the use of algorithms and representations that allow mobile robots to evolve a shared set of symbols where the meanings of the symbols are derived from the robots' sensors and cognition. Mobile robots have two important properties
more » ... that affect the learning of symbols, i) that they are capable of locomotion through space over time; and ii) that they come in many different configurations with different architectures. Previous work has demonstrated that mobile robots can learn shared lexicons to describe space through perceptual referents and referents grounded in cognitive maps. However, open questions remain as to how mobile robots can learn to communicate using temporal terms, and how learning lexicons is affected by different cognitive architectures. The major research question addressed in this thesis is how can mobile robots develop spatial and temporal lexicons across different cognitive architectures? Three facets of language learning are considered particularly important for robots with different cognitive architectures: i) the ability to ground terms in cognition; ii) the ability to ground identical terms in different sensors and cognition for each robot; and iii) the ability to handle referential uncertainty -the difficulty of linking words to meanings within ambiguous contexts. Pairs of mobile robots are used to develop lexicons for spatial and temporal terms and to study each of these abilities. The terms developed by the robots are tested by organizing spatial and temporal tasks and extended to additional terms through grounding transfer. In this thesis, language learning is studied within a framework defined by Peirce's semiotic triangle and building on previous Lingodroid studies. Conversations between robots are used to socially ground symbols within the robots' spatial and temporal cognition. Distributed lexicon tables are used to store links between words and meanings. As the lexicons evolve the words are analyzed for immediate usability, and the final lexicons are analyzed for coherence. Four studies to analyze different aspects of lexicon learning were completed. Study I addressed the aims of learning duration terms using mobile robots and using grounded spatial and temporal language together to perform joint tasks. Identical mobile robots were used to ground terms for time in durations using clocks (time since the last meeting). The robots were able to develop coherent lexicons, and successfully organize future meetings using learned terms. Study II addressed the aim of learning event-based temporal terms using mobile robots. Identical mobile robots were used to ground terms for time in sunlight levels (time of day). The robots required the ability to ground terms in features formed from a brightness level and its derivative. Again the robots were i ii able to develop coherent lexicons and organize meetings, handling changing daylight cycles throughout a year. Study III addressed the aim of learning spatial terms across different cognitive architectures. Robots with different sensors and spatial cognition were used to ground spatial terms within their different spatial representations. These spatial terms could then be used to bootstrap terms for distances and directions, unifying the two robots' different spatial terms into identically represented higher-level terms. The robots were able to develop coherent lexicons for distances and directions. This suggests that the underlying spatial terms -grounded in different spatial sensors and cognition -were also coherent. Study IV addressed the aim of resolving uncertainty using cross-situational learning. The same pair of robots within a simulator were used to ground terms for space and time but in uncertain conditions where the feature of interest was not communicated a priori. The robots in this study used information metrics with cross-situational learning to decide when to link a word and meaning. Cross-situational learning was compared to the lexicon learning from the previous studies on learning time and usability. Results showed that the robots were capable of learning coherent lexicons despite the uncertainty, although with an increase in learning time and a decrease in immediate usability. From the four completed studies, three major conclusions have been drawn. Firstly, the coherence of the lexicons in each study demonstrate that it is possible to i) learn terms for durations grounded in clock time; ii) learn terms for times of day grounded in sunlight levels; iii) ground distances and directions in different underlying spatial representations; and iv) ground spatial and temporal lexicons across different cognitive architectures and achieve communicative success. The second major conclusion is the set of changes to the Lingodroids framework that are required for handling different learning tasks. For a system such as Lingodroids, the ability to generalize the core of the framework over multiple scenarios is one of the most important characteristics. This thesis demonstrates that the same distributed lexicon tables, conversations, categorization and generalization can be used across all study conditions. However, certain aspects of the Lingodroids framework do not generalize, and these aspects represent observations about the key differences between each of the learning conditions. The required changes include referents for new representations of time and space, cross-situational learning, and temporal cognition. The third major conclusion is an expansion of the nature of semiotics. Traditionally a symbol was linked to a referent in the environment through a private representation. However, for robots with different cognitive architectures, a shared symbol may be linked to multiple different private representations, and to multiple different sensors before linking back to a referent in the environment. iii iv Publications during candidature
doi:10.14264/uql.2016.65 fatcat:kkkskjemvvckbfxvn3ads7kzfi