THIS PROJECT IS OBSOLETE! USE POGAMUT 3 INSTEAD

Pogamut : Episodic memory for virtual agent
Print

Episodic memory for virtual agent

The aim of the project

The goal of this work is to design and implement a prototype of the episodic memory for virtual humans. The memory is inspired by up to date research on function of human memory for personal events (episodes) and human time perception. We design a model of memory based on this theoretical knowledge. We took as the point of departure episodic memory system and decision making system of Peskova. The decision making system is based on the BDI, theory of affordances and AND-OR trees. The former episodic memory suffered deficiencies in the recall for time-cued questions. Thus proposed model is working with a unique subsystem for the time perception which allows for more realistic storage and recall of past events. The agent enhanced by this model can reply to questions like “What did you do last week afternoons?”. The prototype is programmed in Java using the framework Pogamut 2. Pogamut 2 is connected to the complex continuous 3D world of Unreal Tournament 2004 which allows us to verify the design in the challenging environment. We have conducted several experiments. The results show that the model extends agents cognitive abilities with a capability to understand socially established temporal patterns. That allows him to answer to the questions with a vaguely specified time information. Moreover, the memory has a limited capability to blend similar episodes together.

Downloads

  • Tutorial (video) 13MB a video tutorial to give you a notion about the project (subtitled in English, no sound).


Short outline of the design

The episodic memory is divided into two main modules – item memory and neural memory.

Item memory

Agent is performing a lots of various tasks and most of those demand one or more resources (prerequisites) before they can be performed. A resource can be in general anything and agent is constantly facing the localization problem when obtaining resources for his actions. Therefore agent needs a representation of space or some memory for resources. This work is focused on the episodic memory thus we will model the memory for items in scarce and simple fashion.
The item memory is storing every item, place and agent he has met. It stores an additional information with each record about the last time he saw it, how many times he found it and how many times he failed to find it on the location. Those information are then used in the simple formula to determine the credibility of the record.

Connectionist memory

The connectionist memory is much more complicated. In short, it is composed from three layers of “neurons” which are interconnected by weighted links (fig. 1). First layer consists of context nodes which describe agent’s internal or external state (context) and Cartesian nodes which represent time or biorhythms (their activation changes regularly over time). First layer is connected with the layer of time concepts. Concept nodes are associating together particular context of agent with some period of time thus representing time concepts. Agent learns them automatically during the simulation. They provide him a vaguer notion of the time. They allows him to answer questions which contain for instance a part of a day (e. g. morning) as the time information.
The next layer of nodes consists of AND-OR trees which are linked together with objects (items, places) and time nodes – either concept nodes, either nodes for days, weeks, etc.



Fig. 1. Episodic memory design outline. The context layer (top left) and Cartesian layer (top right) are connected with concept nodes (center). Concept nodes are connected with AND-OR trees of desires (at the bottom).


Bibliography

  • Nuxoll, A. (2007): Enhancing Intelligent Agents with Episodic Memory. Ph.D. diss. The University of Michigan.
  • Peskova, K. (2006): Model paměti pro animata. Bachelor thesis. The Charles University. Prague (in Czech).
  • Friedman, W.J. (1993): Memory for the Time of Past Events. In Psychol Bull 113(1): 44 – 66.
  • Dayan, P., Abbott, L.F. (2000): Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems. Cambridge: MIT Press, 2001, Chapter 8.
  • Larsen, S.F., Thompson, C.P., Hansen T. (1995): Time in Autobiographical Memory, In: Remembering Our Past: Studies in Autobiographical Memory, New York: Cambridge University Press, 129-156.


Created by: ondrej.burkert. Last Modification: Saturday 08 of August, 2009 11:51:25 CEST by michal.bida.

News

Tiki RSS feed for weblogs

This work is supported by GA UK 1053/2007/A-INF/MFF (2007-8), GA UK 351/2006/A-INF/MFF (2006-8), the Ministry of Education of the Czech Republic (grant MSM0021620838) (2008-9), by the Program "Information Society" under project 1ET100300517 (2006-9), and the project Integration of IT Tools into Education of Humanities (2006-8) and by the project CZ.2.17/3.1.00/31162, which are financed by the European Social Fund, the state budget of the Czech Republic, and by the budget of Municipal House Prague.