Abstract
The overall aim of our work is to develop a generic adaptive control
architecture for autonomous intelligent agents. In our previous work we showed
how a single agent's survival can benefit from autobiographic memory. In the
current work we extend this work by introducing a multi-agent context. Also, we
investigate autobiographic agents which are able to share episodic memories
(sequences of perceptions and actions) in the form of 'communications' with others
in order to find necessary resources for 'surviving' in an environment. This Artificial
Life perspective allows assessment of the potential benefits of such communications
in a multi-agent context. A virtual experimental-based approach deals with different
conditions on agent communications. Extending our previous work, we introduce an
improved control architectures design for autobiographic agents that allows them to
communicate episodic memories. Using detailed measurements of the agents'
lifespan, this is compared both with groups of purely reactive agents and
non-communicative autobiographic agents. Results confirm our previous research
hypothesis that autobiographic memory can prove beneficial, indicating increases in
the lifespan of an autonomous, autobiographic, minimal agent. The paper presents
results from simulation runs varying the following parameters: a) with/without
communications of episodic memories, b) varying the number of agents in the
environment from one to five, and c) communications under different types of
conditions, including presence/absence of costs and motivations. Results show that
in experiments with certain numbers of agents, communicating memories with
others can compensate for the negative effect of generally decreasing lifespan due to
multiple agent interference. Experiments on specific combinations of motivation
and cost conditions also show that, when cost is incurred for communication, agents
perform better if they communicate memories only when they are in need of a
resource.
architecture for autonomous intelligent agents. In our previous work we showed
how a single agent's survival can benefit from autobiographic memory. In the
current work we extend this work by introducing a multi-agent context. Also, we
investigate autobiographic agents which are able to share episodic memories
(sequences of perceptions and actions) in the form of 'communications' with others
in order to find necessary resources for 'surviving' in an environment. This Artificial
Life perspective allows assessment of the potential benefits of such communications
in a multi-agent context. A virtual experimental-based approach deals with different
conditions on agent communications. Extending our previous work, we introduce an
improved control architectures design for autobiographic agents that allows them to
communicate episodic memories. Using detailed measurements of the agents'
lifespan, this is compared both with groups of purely reactive agents and
non-communicative autobiographic agents. Results confirm our previous research
hypothesis that autobiographic memory can prove beneficial, indicating increases in
the lifespan of an autonomous, autobiographic, minimal agent. The paper presents
results from simulation runs varying the following parameters: a) with/without
communications of episodic memories, b) varying the number of agents in the
environment from one to five, and c) communications under different types of
conditions, including presence/absence of costs and motivations. Results show that
in experiments with certain numbers of agents, communicating memories with
others can compensate for the negative effect of generally decreasing lifespan due to
multiple agent interference. Experiments on specific combinations of motivation
and cost conditions also show that, when cost is incurred for communication, agents
perform better if they communicate memories only when they are in need of a
resource.
| Original language | English |
|---|---|
| Title of host publication | 8th Conf on Intelligent Autonomous Systems |
| Subtitle of host publication | IAS-8 |
| Publisher | IOS Press |
| Pages | 361-370 |
| Publication status | Published - 2004 |