Session 1: Capabilities
Goals:
- Completeness/Breadth
of coverage
- Support
any potentially useful primitives
- Architectural
commitments to constraints for
- Efficiency
- Making
“the things you want to do” easier and everything else harder
- To
fit human data
- Valid cognitive
architecture
- Useful
intelligent agent architecture
Current deficiencies:
- Learning
- Category/semantic
- Reinforcement
- Episodic
- Predictive
learning
- Ubiquitous
automated learning
- Speculative
execution
- Representations
- Iconic
- Support
for goal trees
- Planning
mechanisms/representations
- Subcognitive
processing
- Emotion/mood/affect
(human performance moderators)
- Physiological
models
- Activation-based
memory
- Non-symbolic
decision making
- Architectural
clock
- Chunk
decay
- Cognitive
processing
- Managing
deadlocks and resource conflicts
- Standardized
perceptual/motor modules
- Standard
variety of memory buffers
- Partial
matching
- Analogy
- Attention
i.
Cognitive
ii.
Sensorimotor
- Structure
and search of semantic memory
- Bounded
matching, instantiation, and elaboration
- Passing
of time (urgency, time=cost)
- Spatial
situation, reasoning, and representation
- Models
of micro-worlds (as opposed to problem spaces)
- Metacognitive
processing
- Knowledge
revision: can change rules
- Rules
can match against rules
- Arbitrary
function calls in rule conditions
- Forgetting
in long-term memory
- Preferences
and working memory
Session 2: Architecture development usability: extensibility
& embedability
Goals:
- Tease
apart from TCL
- Default
user interface that’s simpler than TCL
- Managing
I/O better (don’t’ crash…easy to do…standard representations)
- Decrease
version specificity
- More
modular kernel
- Cleaner
API to kernel
- API to
kernel components
- Interfaces
to higher-level applications
- Thread-safe/multi-threading
- Support
for multi-agent applications
- Support
for psychologists
- Support
easy/small experimentation with architecture
- Improve
ease of understanding of kernel
- Support
variations on kernel
- Make
sure kernel never hangs
- Ability
to run days at a time
- Ability
to scale to large systems without changing things
Current deficiencies
- Agent
I/O systems
- Standard
functions and data structures for processing I/O
- Reading
and writing XML for I/O (translation to and from WMEs)
- Interfaces
to perceptual motor modules
- Stress
testing of I/O systems (so they don’t crash)
- Kernel
to embedded systems
- Create
an external kernel API that won’t change for at least 10 years
- Create
library of interfaces to other languages and systems
- Do
we have to worry about OS level mechanisms (like paging)?
- Kernel
to multi-agent systems
- Refactor
and separate multi-agent controller (controller distinct from kernel)
- Inter-communication
of agent statistics
- Kernel
to end-users and model developers
- Design
and build a stripped-down user interface
- User-defined
collection of performance and behavior statistics and observations
- Make
available tunable run-time parameters for behavior and performance
- Dynamically
loadable behavior and performance modules
- Callbacks for user variations on
behavior
- Kernel
internals
- Refactor
the kernel
- Eliminate
memory leaks
- New
component-oriented design of kernel
- Well
defined API for kernel components
- Reimplement
in an object-oriented language
- Version
control (CVS)
- Docs
and specs
- A
real BNF for Soar rule syntax
- XML
or Skeptic specification of kernel processing