|PhD Candidate||Office:||Axon Building (47) 307|
|School of Information Technology and||Phone:||+614 0810 1669|
|Electrical Engineering||Email:||scott (dot) heath (at) uqconnect (dot) edu (dot) au|
|The University of Queensland|
|Queensland 4072 Australia|
I am a PhD candidate at the University of Queensland looking at "Effective communication between agents with different cognitive capabilities." I am part of the Lingodroids project.
Lingodroids is a project investigating mobile robots evolving their own language for space and time. Lingodroid agents are based on a symbol grounding methodology. An agent associates a word through a two step process: i) private grounding and ii) social grounding.
Private grounding is forming an internal representation from referents in the agents' environments. The Lingodroid agents, which focus on learning spatial terms, use RatSLAM as their spatial representations.
Social grounding is two or more agents agreeing on a symbol to describe their respective internal representations. A key challenge for this step is sharing the multiple agents' attention to a common referent. Shared attention in Lingodroids is achieved through shared experience. Two agents can share the experience of being at the same place at the same time allowing the agents to name the place and time.
Above is a simulation of two iRats playing language games and developing words for space and time. The top-left square is the two robots in their environment - in previous studies we have used a U and a Q cut from sheet foam. The bottom two squares represent the agents' spatial terms. The red iRat's lexicon on the left and the blue iRat's lexicon on the right. In the top-right square are the two agents terms for time. The red iRat's temporal lexicon on top and the blue iRat's temporal lexicon on the bottom.
A key feature of the Lingodroids methodology is that it makes a distinction between comprehension and production - that is the difference between attempting to understand a word, and choosing a word to describe a referent. In particular generalisation (applying a previously used word to new referents) is only performed during production. In the simulation above, the spatial blobs represent the word that would be produced to describe that location. The regions for comprehension include the regions for production but may also overlap with different words.