Memory Management With Explicit Time in Resource-Bounded Agents

Valentina PitoniAAAI 2018 · DOI

Introduction

In the field of Intelligent Software Agents, memory is frequently modeled as a static repository of facts. However, real-world reasoning is inherently tied to the passage of time. Most existing logical frameworks for "resource-bounded agents"—those with limited processing power or memory—treat beliefs and rules as time-independent. This creates a significant gap: real-world agents must understand that a perception may only be relevant in the present (e.g., "the temperature is low") or that a specific rule has a strict expiration date (e.g., "you must pay this bill by Friday").

To address this, this research introduces TDLEK (Temporalized Dynamic Logic of Explicit Beliefs and Knowledge). By integrating Metric Temporal Logic (MTL) into existing epistemic frameworks, the author enables agents to track when a belief was formed and how long a specific rule remains valid, allowing for more realistic "forgetting" and belief updates.

The TDLEK Framework

TDLEK is an extension of the existing logic known as DLEK (Dynamic Logic of Explicit Beliefs and Knowledge), enhanced with elements of Metric Temporal Logic. This framework allows for a more nuanced representation of an agent's mental state by incorporating explicit time-stamps and intervals.

Key Technical Components

  • Timed Formulas ($\Phi_I$): Atomic propositions are "time-stamped." For example, $p_t$ signifies that proposition $p$ is true at time $t$.
  • Interval-Based Rules: Long-term memory rules are associated with specific time intervals $[t_1, t_2]$ during which they are considered valid.
  • Mental Operations: The logic defines four specific "conscious" operations an agent can perform to manage its internal state:
  • Perception ($+\phi_{I_1}$): Forming a new belief based on a timed observation.
  • Conjunction ($\cap$): Combining two existing timed beliefs into a single coherent thought.
  • Inference ($\vdash$): Deriving a new timed belief based on a rule stored in long-term memory.
  • Update/Restructuring: A specialized operation where a new perception can "overwrite" or shorten the validity of an older belief if they are found to conflict.

Cognitive Architecture: STM vs. LTM

The TDLEK model mimics human cognitive structures by distinguishing between short-term and long-term memory. Short-term memory holds time-stamped perceptions and immediate beliefs, while long-term memory holds the "background knowledge"—the rules that govern how those perceptions are processed.

A crucial element of this architecture is the "Always" Operator ($\Box_I$). This temporal operator signifies that a piece of knowledge is valid throughout a specific interval, which is essential for representing legal regulations, scheduled tasks, or persistent environmental states.

Belief Defeasibility and Restructuring

One of the primary advantages of TDLEK is its handling of defeasibility—the ability to retract or update a belief when new, more timely information arrives. Unlike simple "forgetting," which merely deletes data, TDLEK allows for "restructuring."

For example, if an agent believes they are enrolled in a university for the duration of a year, but then takes a full-time job at month eight, the logic provides a formal way to truncate the "enrollment" belief so that it ends at month eight. This ensures the agent's internal model of the world remains consistent with the actual timeline of events.

Case Study: University Enrollment

The primary result of this work is the formal definition of the TDLEK language and its semantics. The author demonstrates the utility of the logic through a University Enrollment scenario:

  • Rule Processing: The agent uses rules to understand that passing a high school exam at time $t_0$ opens a specific window for university enrollment.
  • Dependency Tracking: The agent can track complex temporal dependencies, such as: "If I register between day 213 and 304, I must pay the fee by day 334."
  • Conflict Resolution: The logic successfully handles conflicting information, such as an agent quitting university early, by updating the time intervals of its active beliefs to reflect the change in status.

Conclusion and Future Work

TDLEK adds significant power to the practical expressiveness of agent logic, allowing resource-bounded agents to interact with dynamic environments in a timely and logical manner. However, the author acknowledges that challenges remain. While the logic is designed for resource-bounded agents, further work is needed to fully overcome the computational challenges of reasoning with strictly limited resources. Future research will focus on refining how agents "restructure" memory to ensure they do not become overwhelmed by historical data that is no longer relevant to their current goals.

Summary updated

BibTeX
@inproceedings{pitoni2018memory,
  author = {Valentina Pitoni},
  title = {{Memory Management With Explicit Time in Resource-Bounded Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.12136},
}
Figure