Human-Inspired Forgetting for Robotic Systems
Freedman, Sanford Tory
--PLEASE NOTE THAT THE ATTACHED .7z FILE CONTAINING VIDEOS WILL NEED TO BE EXTRACTED AND BURNED TO A DVD IN ORDER TO BE VIEWED.-- Perfect memory and recall provides a mixed blessing. While flawless recollection of episodic data and procedural rules allows for increased reasoning, photographic memory hinders a robot's ability to operate in real-time, highly dynamic environments. The absence of forgetting can result in memory being filled by a tremendous volume of data, increasing both search time and the probability of over-learning. Many small, but critical details within the environment greatly impact the probability of successful task completion, unfortunately robots are currently ill-equipped to navigate incoming data to detect, recognize, and act upon these details. As robotic hardware and designs improve, robots will be further inundated as finer resolution environmental data and higher accuracy mental models become available. Contemporary robots are already overrun with vast volumes of data requiring real-time processing and the problem will only increase. Before robots realize human-level intelligence, a means of classifying the importance of each acquired datum and forgetting unnecessary, erroneous, and expired data will be required. This dissertation has developed Human-Inspired Forgetting, a means of incorporating forgetting capabilities into current and future robotic systems. This approach may enable robots to remove unnecessary, erroneous, and out-of-date information while increasing the ability to reliably and rapidly recall critical cues necessary critical for successful task completion. Instead of selecting an item from memory to complete a task, Human-Inspired Forgetting filters the information presented to existing robotic algorithms. The pruned data may allow a diverse array of powerful, but task specific algorithms to realize improved accuracy while reducing cognitive load. The novel ActSimple forgetting algorithm has been developed as an implementation of Human-Inspired Forgetting. This forgetting algorithm has been heavily inspired by a number of cognitive architectures along with models of human memory and incorporates trace-based decay, encoding interference, belief state values, mental exertion, and output interference. Simulation and real world experiments were conducted to demonstrate the performance and reliability of Human-Inspired Forgetting and the ActSimple algorithm across a range of testing conditions.