At RulesFest 2011, I listened to Nate Derbinsky on “Effective Scaling of Long-term Memory for Reactive Rule-based Agents.” Nate is a post-doc researcher working with John Laird and Soar. Perfect talk for the end of the day… Playing with robots… Plus, there were cool references to Memento and Star Trek!
In his example, “iRobot” moves toward a cliff edge and encounters a stop sign in one scenario and an explosive device in another. According to Nate, intelligent agents need effective access to diverse information to behave appropriately. However, traditional rule-based and search-based approaches take too long. His research focuses on real-time reactivity (less than 50 msec).
Agents, like humans, rely on both semantic and episodic memories to know how to respond in a given scenario:
- Semantic memory is factual. It’s how you know something.
- Episodic memory, on the other hand, is experiential. It’s how you remember something. Episodic memory is what traditional approaches lack.
To understand his architecture design for both, I recommend going through his slides as they’re quite detailed AND animated. A picture is definitely worth a thousand words!
In addition to robots, Nate Derbinsky and team also tested episodic memory in game, planning domain definition language (PDDL), and word-sense disambiguation (WSD). Super exciting!
Learn more about Decision management and Sparkling Logic’s SMARTS™ Data-Powered Decision Manager
See our recap on Davide Sottara’s talk