Unit 7: Game Playing: Overview

Topic: The Blocks World in AI (e.g. game of 3 glass, in

which has ball)

The Blocks World is a classic planning problem in artificial intelligence. It

provides a simple and intuitive domain to test and demonstrate various planning
techniques.

Scenario

 The world consists of:

o A table
o A set of blocks (usually labeled A, B, C, etc.)
o A robot arm that can manipulate the blocks
 Each block can either be on the table, on another block, or held by the
robot arm.

Typical Assumptions

 Only one block can be moved at a time.

 A block can only be moved if there is nothing on top of it.
 The robot arm can hold only one block at a time.
 Blocks can be stacked or placed on the table.

Initial State & Goal State

 The problem is defined by:

o An initial state: the current arrangement of the blocks
o A goal state: the desired arrangement
 The planner must find a sequence of actions to reach the goal from the
initial state.

Actions (Operators)

Each action has preconditions and effects:

1. Pickup(block)
o Preconditions: block is clear, block is on the table, arm is empty
o Effects: block is held; block is no longer on table
2. Putdown(block)
o Preconditions: arm is holding block
o Effects: block is on table, block is clear, arm is empty
 3. Stack (block1, block2)
o Preconditions: block1 is held, block2 is clear
o Effects: block1 is on block2, block1 is clear, arm is empty
4. Unstack (block1, block2)
o Preconditions: block1 is on block2, block1 is clear, arm is empty
o Effects: block1 is held, block2 is clear

Why It’s Important

 Simple to understand but reveals complex planning issues like:

o Sub goal interactions (e.g., achieving one goal undoes another)
o Nonlinear planning (not all steps must be in strict sequence)
o Goal ordering (figuring out the best order to achieve goals)
o State representation and action preconditions/effects

Use in AI

 Used in:
o STRIPS (Stanford Research Institute Problem Solver) planning
systems
o Goal Stack Planning
o Hierarchical Task Networks (HTN)
o Partial-order planners
 Serves as a benchmark for comparing planning algorithms.

Example Problem

Initial State:

 A is on B
 B is on the table
 C is on the table
 Arm is empty

Goal State:

 C is on A
 A is on B
 B is on the table

So, answer: A planner would have to:

1. Unstack A from B
2. Put A on the table
 3. Stack C on A

Topic: components of planning system in game playing in

ai (e.g. level type any game)
1. State Representation

 Describes the current configuration of the game.

 Includes:
o Positions of pieces (in chess, tic-tac-toe, etc.)
o Whose turn it is
o Any relevant game-specific data (score, remaining time, etc.)
 Important for evaluating future actions.

2. Action Generator (Move Generator)

 Lists all legal moves (actions) available from the current state.
 Must obey the rules of the game.
 Example: In chess, this would generate all possible valid moves for a
player.

3. Transition Model (Successor Function)

 Defines how the state changes when an action is performed.

 Applies a move and returns the resulting new game state.
 Essential for simulating future moves (used in MiniMax, etc.).

4. Goal Test / Terminal State Checker

 Determines whether the game has reached a winning, losing, or draw

condition.
 Helps stop the search at terminal nodes in the game tree.

5. Evaluation Function (Heuristic Function)

 Estimates how desirable a non-terminal state is.

 Used when it's impractical to search all the way to the end of the game.
 Example:
o In chess: material balance + positional advantages
o In tic-tac-toe: number of winning lines open

6. Search/Planning Algorithm
  Chooses the best move by searching through future states.
 Examples:
o MiniMax
o Alpha-Beta pruning
o Monte Carlo Tree Search
o Iterative Deepening
 Tries to maximize the player’s advantage while minimizing the
opponent’s.

7. Time Management (Optional but Important)

 Especially in real-time games or games with a time limit.

 Allocates time for move computation using:
o Iterative deepening
o Time-based cutoffs
o Adaptive depth search

8. Memory / Transposition Table (Optional Enhancement)

 Stores previously visited states to avoid re-evaluation.

 Speeds up search by recognizing duplicate states (e.g., in chess, many
move sequences can result in the same board).

Summary Table:
Topic: Goal State in Planning (AI)
What is a Goal State?

 A goal state is the desired outcome or configuration that an AI agent

aims to achieve.
 It represents the end condition of a successful plan.

In Simple Terms:

The goal state is like the finish line in a race — it's where you want to end up.

Example 1: Blocks World

 Initial State: A is on B, B is on the table, C is on the table

 Goal State: C is on A, A is on B, B is on the table

Here, the planner must find a series of actions to transform the initial state into
the goal state.

Example 2: Game Playing (like Chess)

 Goal State for a player could be:

o Checkmating the opponent (in chess)
o Aligning 3 Xs or Os (in tic-tac-toe)
o Reaching a specific score or condition

Types of Goal States

1. Explicit Goal State

o Clearly defined final arrangement.
o Example: "Block A must be on Block B."
2. Partial Goal State
o Only some parts of the world need to be in a specific configuration.
o Example: "Block A must be on the table" (no mention of other
blocks).
3. Goal Test (Function)
o A Boolean function that returns true if the current state satisfies
the goal.
o Example: is goal(state) → returns True if checkmate is achieved.
Importance in Planning

Planning in game playing is a fundamental aspect of Artificial Intelligence

(AI) because it enables an agent (like a player or computer-controlled
character) to make intelligent decisions over time.

Planning in AI game playing is crucial for:

 Intelligent decision-making,
 Long-term strategy,
 Handling complex, dynamic environments,
 Creating believable and capable opponents.

How AI Uses Goal States

 In search algorithms: used to stop the search when the goal is reached.
 In heuristics: used to measure how "far" a current state is from the goal.
 In constraint satisfaction: used to define what conditions must hold.

Topic: nonlinear planning using constraints

posting in ai
What is Nonlinear Planning?

 Nonlinear Planning (also called partial-order planning) allows actions

to be partially ordered, not strictly sequential.
 It enables the planner to identify independent actions that can happen in
any order or even in parallel.
 Based on the principle of least commitment: don’t decide the order of
actions unless absolutely necessary.

What is Constraint Posting?

 In constraint-based planning, the planner posts constraints instead of

generating full action sequences.
 Constraints define relationships between actions:
o Ordering constraints (A before B)
o Causal links (A produces a condition required by B)
o Threats (some action may interfere with a causal link)
Core Concepts
Term Description
Action Has preconditions and effects
Causal Link A →p→ B (Action A provides p, which B needs)
Ordering
A < B (A must happen before B)
Constraint
An action C might destroy the condition p between A and
Threat
B
Threat Resolution Handle threats by promotion or demotion

Planning Process (Partial-Order Planning with Constraint Posting)

1. Initialize Plan:
o Add a Start and Finish action.
o Start: has initial conditions
o Finish: has goal conditions
2. Identify Open Preconditions:
o Preconditions of actions that are not yet supported.
3. Post Causal Links:
o Choose an action that achieves the precondition.
o Create a causal link: Achiever →p→ Needer
4. Add Ordering Constraints:
o Ensure the achiever happens before the needer.
5. Check for Threats:
o If any action might delete p between A and B, resolve it:
 Promotion: place the threat after B
 Demotion: place the threat before A
6. Repeat until:
o All preconditions are satisfied
o No threats remain
o The plan is consistent and complete

Example (Simplified)

Goal: Stack A on B, B on C

Steps:

 Add actions:
o Stack(A, B)
o Stack(B, C)
 Post causal links:
o Clear(B) → Stack(A, B)
 o Clear(C) → Stack(B, C)
 Add ordering constraints based on dependencies.
 Check for any actions that interfere with these links and resolve.

Advantages

 Flexibility: Only imposes necessary ordering.

 Parallelism: Actions that don’t interfere can be performed in parallel.
 Scalability: More efficient for large and complex domains.
 Modularity: Easy to modify by adding or adjusting constraints.

Used In

 POP (Partial Order Planner)

 Hierarchical planners
 Automated scheduling and robotics

Topic: Hierarchical Planning in Game Playing (AI)

What is Hierarchical Planning?

Hierarchical Planning, also known as Hierarchical Task Network (HTN)

planning, is a method where:

 High-level goals are decomposed into simpler sub-tasks.

 Planning is done at multiple levels of abstraction.

Key Idea

“Don’t plan everything from scratch. Start big, then break it down.”

For example:

 High-level task: Win the game

 Sub-tasks:
o Build army
o Gather resources
o Attack enemy base

Eac h of these can be further broken down:

 Build army → Train soldiers, Build barracks

 Train soldiers → Collect gold, Assign unit
In Game Playing: How It Works

Hierarchical planning helps AI:

1. Think Strategically: Focus on high-level goals first

2. Break Down Complex Goals: Turn goals into manageable steps
3. Reuse Plans: Reusable templates for common situations
4. React Faster: Less decision-making needed at low levels

Example: Real-Time Strategy Game (e.g., StarCraft)

High-Level Goal Sub-Tasks

Destroy enemy base Build army → Scout enemy → Launch attack
Build army Build barracks → Train units
Train units Collect resources → Select unit type

Each task can be handled by specialized sub-planners.

HTN Planning Components

Component Description
Tasks High-level goals (can be abstract or primitive)
Methods Ways to decompose tasks into sub-tasks
Operators Actual executable actions (lowest level)
Planner Chooses and applies methods recursively

How It’s Different from Classical Planning

Classical Planning Hierarchical Planning

Flat list of actions Multi-level abstraction
Works with atomic actions Works with tasks and sub-tasks
Search is at one level Search is at multiple levels
Less scalable for big problems More scalable and intuitive
Advantages in Game AI
  ✅ More human-like planning
 ✅ Reusability of strategies
 ✅ Better management of complex, multi-step tasks
 ✅ Supports modular AI design (e.g., strategy planner, combat planner,
builder)

Example: Chess AI (Simple View)

Goal: Win the game

→ Control centre of the board
→ Develop pieces
→ Protect king

Each can further be split into:

 Move knight to f3
 Castle kingside
 Exchange pieces

Real-World Usage

 Video game AI (e.g., RPGs, strategy games)

 Robotics
 Simulation and training systems

Summary Table
Element Role in Hierarchical Planning
Task High-level goal (e.g., Win game)
Method Plan to achieve task (e.g., Build army)
Sub-task Step in method (e.g., Build barracks)
Operator Executable action (e.g., Place building)

Topic: Reactive Systems in Game Playing (AI)

What is a Reactive System?

A Reactive System is an AI system that:

  Responds immediately to changes in the environment (like the game
world).
 Has no long-term planning — just fast, rule-based responses.
 Works well in dynamic and real-time environments.

Reactive vs. Deliberative AI

Feature Reactive System Deliberative System

Planning No Yes
Response Time Instant / Fast Slower
Complexity Low High
Memory of Past Minimal or None Maintains state/history
Adaptability Limited High

Key Characteristics

 Uses simple rules or behaviors (e.g., IF-THEN rules).

 Doesn’t build full models of the environment.
 Focuses on the current situation only.
 Often used in enemy AI, non-player characters (NPCs), or agents in
fast games.

Example in Game Playing

e.g. Enemy NPC in a Shooting Game:

 Reactive Behavior:
o IF player is near → Attack
o IF health is low → Run away
o IF sees power-up → Go to power-up

This AI doesn’t plan out strategies or future steps — it reacts based on the
present state.

Reactive System Architecture

1. Sensors (Perception):
o Detect environment (e.g., location of player, obstacles)
2. Behavior Rules:
 Set of conditions and actions (IF-THEN)
o
3. Actuators:
o Execute action (e.g., move, attack, hide)

Think of it like a robot or character running on reflexes.

Types of Reactive Architectures

 Finite State Machines (FSMs):

o Simple models with states like Idle, Attack, Flee
 Subsumption Architecture:
o Layers of behaviour (e.g., "Avoid obstacles" overrides "Wander")
 Behaviour Trees:
o Hierarchical control structure used in many modern games

Pros and Cons

Pros Cons
Fast response No strategic planning
Easy to implement Can behave unpredictably
Suitable for real-time environments Limited intelligence and memory
Scalable for simple agents Doesn’t learn or adapt well

Real-World Game Examples

 Pac-Man Ghosts: React to player’s position.

 FPS Bots: Use reactive strategies for shooting/hiding.
 Racing Games: Opponents react to collisions or track layout.

Summary

Feature Reactive System in Game AI

Planning None
Uses Real-time actions, NPC behaviors
Structure Rules, states, or behavior trees
 Feature Reactive System in Game AI
Response Type Instant, based on current input
Common Use FPS, RPG, RTS, arcade games