Agent Properties
Agent properties determine how suited an agent is for
various tasks. In this document, properties are differentiated
from capabilities (e.g. learning, planning, etc.) by being
more fine-grained. Different architectures may exhibit
a capability to "learn", but may do so at different times
for different types of data. These fine distinctions may
be explored below.
- Declarative vs Procedural:
How is the knowledge stored and used? Declarative knowledge is more flexible
because it may be reasoned about, procedural knowledge is faster to use and
may be cheaper to store.
- Global representation vs no Global Representation:
Do they have global knowledge? Global knowledge is necessary for representing
a unified world-view, but maintaining this world-view is problematic.
- Symbolism vs Subsymbolism:
Can they use symbols? Symbols allow flexible, combinatorial representations.
- Black box vs Glass Box control knowledge:
May agent rules inspect other rules? If so then it is easier to reason about
its own behavior (metareasoning). If not then it is less fragile to
changing what it knows.
- Consistency vs inconsistency:
Must all knowledge be consistent? Knowledge consistency may be used to
implicitly store knowledge via the closed-world assumption, but it
limits the reasoning and hypotheses an agent may have.
- Uniform knowledge rep. vs non-uniform knowledge rep.:
Is all the knowledge in an architecture represented uniformly? Uniformity
simplifies module interfaces at the expense of placing bounds on each.
Reasoning-Related Properties:
Learning-Related Properties:
- Deliberative vs Reflexive:
May the architecture decide whether or not to learn something? Not learning a
rule may save space and matching time but it takes space and time to store
statistics to judge rules. The problem space search vs. knowledge search
tradeoff is part of the utility problem.
- Monotonic vs non-monotonic:
Must the architecture learn monotonically? Learning monotonically is easier
and less likely to cause inconsistencies, but it's not suitable for all
environments.
Organization-Related Properties:
- Centrality vs Modularity:
Are they centrally controlled? Centrally controlled agents may be more suited
for some tasks, less so for others.
Performance-Related Properties:
- Eager vs Lazy sensing:
Is sensing considered a cheap operation? In eager sensing up-to-date views of
the world are considered desirable and easy to get, in lazy sensing they may
be desirable, but possible expensive because of the processing, communication,
etc. involved.
- Quick vs slow: Do they have rapid response?
This is required for real-time applications in dynamic environments.
