. Types of Agents and Planning

The lecture begins by categorizing agents and detailing how advanced agents think.

• Four Basic Agent Types: Simple reflex agents, reflex agents with state/model, goal-based agents, and utility-based agents "refer to prev lecture".
  
  
• Planning Agents: These are advanced agents that ask "what if" before acting.
  
  
  • They require a model of how the world evolves in response to their actions.
    
    
  • They base their decisions on the hypothesized consequences of their actions and must formulate a goal to test against.
    
    
• Problem-Solving Agents: These are a specific type of goal-based agent that consider future actions and the desirability of their outcomes.

2. Formulating the Problem

Before an agent can search for a solution, the problem must be strictly defined.

• The Four Phases of Problem Solving: Goal formulation, Problem formulation, Search, and Execution.
  
  

• The Five Components of a Formal Problem:

  • Initial state: Where the agent starts.

  • Possible set of actions: What the agent can do.

  • Transition model: A description of what each action does and its resulting consequence.

  • Goal test: A check to determine whether a given state qualifies as a goal state.

  • Path cost: The measurable cost of taking a specific sequence of actions.

3. Representing the Environment

When designing agents, you have to decide how complex the agent's internal representation of the world needs to be. There are three levels of abstraction:

• Atomic: The state is treated as a simple "black box" with no internal structure. This is typically used in basic search algorithms, game-playing, and Hidden Markov Models. Problem-solving agents use this representation.
  
  
• Factored: A more complex abstraction where states are represented by a set of variables, properties, or a vector of attribute values (e.g., Location: Room A, Battery: 80%). Planning agents often use this.
  
  
• Structured: The most complex representation, where states include explicit relationships between different objects, and actions can affect specific variables or relationships. Planning agents also frequently use this.

4. Graphs vs. Trees

To understand search algorithms, you need to understand the underlying mathematical structures used to model them .

• Graphs: Mathematical structures made of vertices (nodes) connected by edges (arcs).
  
  
  • Nodes represent discrete states of a problem (like a game board configuration).
    
    
  • Arcs represent transitions or legal moves between those states.
    
    
• Trees: A specific type of graph. It is a connected, acyclic (no loops) undirected graph where nodes have a single parent. All trees are graphs, but not all graphs are trees.
  
  
• Conversion: You can turn a graph search problem into a tree search problem by avoiding loops in your path or keeping track of globally visited nodes.

4.2 Search Problem

A search problem is a formal way to define a problem so that an intelligent agent can solve it by searching for a solution.

Components of a Search Problem

A search problem consists of the following elements:

• State Space: A set of all possible states the environment can be in.
• Successor Function: This includes the possible actions available and their associated costs.
• Start State: The initial state where the agent begins.
• Goal Test: A condition used to determine if a given state is a goal state.

Solutions to Search Problems

• The Plan: A solution is defined as a sequence of actions (a plan) that transforms the start state into a goal state.
• Representations: Search problems often use an atomic representation, where each state is treated as a "black box" without internal structure.
• Mathematical Models: These problems are often represented mathematically using state space graphs, where nodes represent states and arcs represent actions or transitions. They can also be visualized as search trees, which map out "what if" scenarios and potential future outcomes.

5. State Space Graphs vs. Search Trees

This is a critical distinction in the lecture regarding how search is actually executed.

• State Space Graphs: This represents the actual layout of the problem.
  
  
  • Nodes are abstracted world configurations.
    
    
  • Crucially, in a state space graph, each state occurs only once.
    
    
  • It is rarely built fully in memory because it is usually too large.
    
    
• Search Trees: This represents the agent's process of finding a solution.
  
  
  • It is a "what if" tree of potential plans and outcomes.
    
    
  • The root node is the start state, and children are successors.
    
    
  • Crucially, each node in a search tree represents an entire path from the start state, not just a single physical location. This means a single state from the State Space Graph can appear multiple times in a Search Tree if there are multiple ways to get there.

6. The Tree Search Process

The lecture concludes by outlining the general strategy for a tree search algorithm.

• The Fringe: As the agent explores potential plans (expanding tree nodes), it maintains a "fringe" of partial plans currently under consideration.
  
  
• The Loop: The algorithm loops by choosing a leaf node from the fringe for expansion based on a specific strategy.
  
  
  • If it's the goal, it returns the solution.
    
    
  • If not, it expands the node, adding the resulting new nodes back to the search tree's fringe.
    
    
• Efficiency: The main goal is to expand as few tree nodes as possible while deciding which fringe nodes to explore.