افزایش عوامل هوشمند با حافظه اپیزودیک

For a human, episodic memory is a memory of past experiences that one gains over a lifetime. While episodic memory appears critical to human function, researchers have done little to explore the potential benefits for an artificially intelligent agent. In this research, we have added a task-independent, episodic memory to a cognitive architecture. To frame the research, we propose that episodic memory supports a set of cognitive capabilities that improve an agent’s ability to sense its environment, reason, and learn. We demonstrate that episodic memory enables agents created with our architecture to employ these cognitive capabilities.

One advantage that humans have over current Artificial Intelligence (AI) systems is a personal history of specific events that they can draw upon to improve their decision making and learning. This episodic memory was first described in depth by Tulving, 1983 and Tulving, 2002. Tulving’s focus was phenomenological and in particular distinguished episodic memory from semantic memory. Episodic memory provides humans with the ability to remember where they have been, what they have sensed, and what actions they have taken in various situations. This knowledge of the past is invaluable for acting in the present. Episodic memory supports a wide range of cognitive capabilities from keeping track of the world outside immediate perception, to allowing retrospective learning on previously encountered situations. Certainly, there is evidence that human cognition is severely crippled by the loss of episodic memory and the difficulties that amnesiacs face have been documented (Tulving, 2002) and were dramatically portrayed in the movie Memento (Nolan, 2000). As in any learning system embedded in a performance system, episodic memory involves: the capturing and encoding of experience in an internal format; storing that experience in a knowledge base for future use; retrieving knowledge in the future when given an appropriate cue. In addition to supporting these fundamental operations, there are additional functional requirements that Tulving (1983) identified that distinguish episodic memory from other memory and learning mechanisms: • Automatic: The system creates new memories automatically without the agent deciding to do so. The underlying assumptions are that: (a) deliberately deciding which situations to remember can interfere with task-based reasoning and (b) it is unlikely that the agent can effectively determine which experiences will be relevant to future decisions. • Autonoetic: A retrieved memory is distinguished from current sensing, so that an agent does not confuse a retrieved memory with the current situation. • Temporally indexed: Because an episodic memory describes a particular, unique moment in time, some temporal information is a part of any episodic memory and can also be part of an episodic memory cue. This need not be an exact time but it should convey a sense of the relative time of the episode with respect to other episodes. This paper presents our progress toward creating a general purpose episodic memory within a cognitive architecture that supports the creation of general AI agents, that is agents that use large bodies of knowledge, continually learn from experiences in their environment, pursue multiple diverse tasks, and exist for extended periods of time. Although there has been sporadic research on episodic memory within AI in the past, there has not been research on task-independent episodic memories that support a wide variety of cognitive capabilities within a cognitive architecture. Thus, our research involves determining the requirements for an episodic memory; designing, implementing, and integrating an episodic memory system within a cognitive architecture; and exploring the capabilities supported by such an integration. The emphasis of our research has been to create a computational system with the most important features of episodic memory so that we can develop and evaluate not just an episodic memory module, but an integration of that module within a cognitive architecture in which we can build agents. This paper describes our progress to date on this work. While we have made considerable progress, our episodic memory architecture is far from complete. For example, it does not include memory consolidation, forgetting, interference, priming, generalization across episodes or specific models of time. Our architecture does support effective encoding, storage and retrieval and we have used it to create agents for a variety of tasks. Our research suggests that episodic memory enhances the performance of AI agents and may be a “missing link” in current cognitive architectures, enabling a gamut of cognitive capabilities

In this paper, we presented a broad picture of the challenges and benefits of providing an episodic memory to an intelligent agent. We then cataloged some of the cognitive capabilities episodic memory might be able to support in an agent. Next, we described an implementation of a general-purpose episodic memory for the Soar architecture. Finally, we presented the results from using our episodic memory architecture’s ability to facilitate these cognitive capabilities. By investigating these capabilities and demonstrating the possibility of supporting them in an artificially intelligent agent, we have established the possibility that a single, general memory system can provide them. This research presents demonstrations of five cognitive capabilities: virtual sensing, action modeling, decision-making based on past experiences, retroactive learning and boosting other learning mechanisms. In addition to those cognitive capabilities, we have identified the following additional cognitive capabilities that intuitively appear to require an episodic memory. Given our goal to create a comprehensive episodic memory, our future work will focus on demonstrating these additional capabilities with our system. • Noticing significant input: One challenge for an agent is determining which aspects of its current situation are most important, with changes to the environment being an important indicator of relevance. Episodic memory allows an agent to detect changes by comparing the current situation to prior memories. • Detecting repetition: Given the limited memory of most AI agents, it is difficult for them to detect when they repeat the same sequence of actions without making any progress on their current task. Episodic memory provides the necessary memory to detect when the same situation is encountered, or the same action is tried repeatedly. • Environment modeling: In many domains, the environment has its own dynamics (e.g., Sunset has been around 6:30 pm lately.). An episodic memory provides a record of these changes and, thus, allows the agent to predict them in similar situations in the future. • Managing long term goals: An agent with multiple goals must often switch between them because of environmental demands and opportunities. This requires that the agent be able to record the progress it has made for a given goal and restart or recover its progress. Furthermore, a goal must sometimes be suspended for an indefinite period of time. An agent can create a prospective episodic memory ( Kliegel et al., 2007), i.e., it can remember to reinstantiate a suspended goal in response to a future expected event. To schedule these goals, an agent needs to recall goals it has suspended and remember the progress it has made toward each one. • Sense of identity: An agent with a sense of identity potentially gains a greater ability to recognize its own behavior and analyze it compared to the behavior of other agents. For humans, one’s sense of identity is rooted in memories of past experiences, which indicates that episodic memory has an important role to play in this capability. • Reanalysis given new knowledge: When a learning agent receives new knowledge about its environment, inferences and behavior it has learned in the past may no longer be valid. An episodic memory allows an agent to review its experiences that relate to the new knowledge and change its behavior accordingly. • Explaining behavior: The ability to remember what you did in the past allows you to explain your actions to others and allow them to instruct you or you to instruct them (e.g., Why did you go left instead of right?). An agent can use its episodic memory to recall the situation in question as well as the decisions it made in that situation. We assess the strengths and weaknesses of our approach to creating a task independent episodic memory as follows. Integrating our memory system into a cognitive architecture means any agent constructed with that architecture automatically gains the capabilities granted by that episodic memory. Furthermore, the episodic memory operates within the restrictions defined by a body of existing research aimed at creating general intelligence. By carefully examining the design decisions that were made in constructing our episodic memory system, we implicitly define a set of possible, alternative implementations. This allows us to methodically compare specific implementations and select the ones that are most effective. However, this space of possible implementations is by no means complete. Design decisions that we overlooked will hide portions of the space from our investigation. By defining, in advance, a set of cognitive capabilities that could be facilitated by an episodic memory, we provide a metric for measuring the quality of an episodic memory architecture. These capabilities also provide a clear direction for future research. However, due to their introspective nature, any set of cognitive capabilities we define is inherently imperfect and likely incomplete. Furthermore, the qualitative nature of some of these cognitive capabilities makes them difficult to measure. Finally, by building a complete episodic memory system and refining it we move expediently to a system that can be used for various research tasks. As a result, we rapidly gain experience and insight about the abilities and limitations of episodic memory. However, by committing to a single approach it is more difficult to gain perspective on the best method for building an episodic memory architecture.