This year the AAAI Fall Symposium Series included a track organized by Pat Langley on Advances in Cognitive Systems. Pat identified the following objective for the symposium:
Pursue the initial goals of artificial intelligence and cognitive science: To explain intelligence in computational terms and reproduce the entire range of human cognitive abilities in computational artifacts.
Pat’s motivation for re-evaluating the goals of artificial intelligence is presented in a recent editorial in Machine Learning, while my motivation for attending the symposium was to discuss how to build integrated, heterogeneous agents. The introduction to the symposium was a discussion of why AI research has gone astray. Pat identified the following issues in his talk:
- Increased speed and storage allow dumb methods to work on limited problems, distracting from work on general intelligence.
- Focus on quantitative metrics leads to incremental work on well-established problems.
- Theorem envy.
- Commercial success on narrow problems.
After discussing how these issues have constrained AI research, Pat proposed the following goals for cognitive systems:
- Focus on high-level cognition
- Structured representations
- Systems perspectives
- Influence on human cognition
- Exploratory research
The main takeaway I got from the the talk was that a portion of AI researchers should be working on building complete systems as well as working on problems such as integrating AI methodologies in agents.
The symposium included a wide range of topics including cognitive architectures, language understanding, qualitative reasoning, and evaluation of cognitive systems. The symposium website contains a complete list of papers and posters presented at the event.
A number of research groups are using the SOAR and ICARUS cognitive architectures to build intelligent systems for diverse domains. Randy Jones discussed how SOAR has been applied to a tactical air simulation, and made the point that if you are going to include new components in your cognitive architecture, then it is necessary to have problem sets which require these components. Dongkyu Choi presented the learning capabilities in ICARUS and explained how learning from failures and learning from observations have been added to the system. Paul Schermerhorn provided an overview of work being done to integrate the DIARC robot execution environment with ICARUS.
During the poster session at the symposium, I presented my paper on integrating machine learning and case-based reasoning components in a reactive planning agent. EISBot plays the real-time strategy game StarCraft at the same granularity as human players. One of the discussion points that came up during the session was that the ABL reactive planning language provide a good framework for performing goal management for real-time agents, while it may be beneficial to interface the system with other architectures which provide mechanisms such as skill learning.
One of the questions that came up during the event was how to evaluate cognitive systems. Given that one of the goals is to build complete systems, it is necessary to move beyond standardized problem sets. But its no longer enough to just show a system working on a previously unexplored problem. For now, I am planning on testing against human participates to evaluation EISBot.
The symposium had an excellent turn out for its first year with 76 registrants, and another event on advances in cognitive systems will be planned for next year. Updates will be posted here.
