Arquitecturas cognitivas para agentes

Welcome! Today, we'll explore Cognitive Architectures. These aren't just simple programs; they're frameworks designed to mimic human-like reasoning, learning, and decision-making in AI agents.

They bridge the gap between reactive agents and truly intelligent systems capable of complex problem-solving.

Simple agents react to immediate percepts. But what if an agent needs to plan, learn from mistakes, or understand complex situations?

• Human-like Intelligence: Model how humans think.
• General Problem Solving: Tackle diverse tasks, not just one.
• Learning & Adaptation: Improve performance over time.
• Robustness: Handle unexpected situations.

Most cognitive architectures share common building blocks, inspired by human psychology:

• Perceptual System: How the agent "sees" the world.
• Memory Systems: Short-term (working) and long-term knowledge.
• Decision-Making: How the agent chooses its next action.
• Motor System: How the agent acts on the world.

SOAR (State, Operator, And Result) is a classic cognitive architecture. It views all intelligence as a continuous process of problem-solving, represented as searching through a state space.

SOAR operates in decision cycles, constantly choosing operators to apply to the current state to reach a desired result.

SOAR's working memory holds the agent's current understanding of the world, its goals, and the current problem state. It's temporary and constantly updated.

Think of it as the agent's "consciousness" at any given moment. Here's a simplified representation:

SOAR uses production rules (if-then rules) in its long-term memory to propose and select operators. When a rule's if condition matches the working memory, its then part proposes an action or modifies the state.

This example shows a simple rule for moving an agent:

A unique feature of SOAR is chunking. When the agent encounters an impasse (a situation where it can't decide what to do), it enters a sub-state to resolve it.

Once the impasse is resolved, SOAR "chunks" the experience, creating a new production rule that directly solves that type of impasse in the future. This is how SOAR learns!

ACT-R (Adaptive Control of Thought—Rational) is another prominent cognitive architecture. It's designed to model human cognition at a finer grain, focusing on psychological data and predicting human behavior.

ACT-R emphasizes a modular structure, with distinct memory systems and processes working together.

ACT-R has several key modules, including:

• Declarative Memory: Stores factual knowledge (e.g., "Paris is the capital of France") as discrete units called chunks.
• Procedural Memory: Stores "how-to" knowledge (e.g., "how to tie a shoe") as production rules.
• Goal Module: Manages the agent's current goals.
• Imaginal Module: Holds temporary problem representations.

Unlike SOAR's pure rule-matching, ACT-R's modules interact based on activation and utility:

• Activation: Chunks in declarative memory have an activation level, influencing how quickly they can be retrieved. More relevant or recent chunks have higher activation.
• Utility: Production rules in procedural memory have a utility value, reflecting their past success. Rules with higher utility are more likely to be chosen.

We've explored Cognitive Architectures, frameworks that aim for human-like intelligence. We looked at:

• SOAR: Problem-solving as search, using production rules and learning via chunking.
• ACT-R: A modular system with declarative and procedural memory, guided by activation and utility.

These architectures provide powerful models for building agents that can reason, learn, and adapt in complex ways.