Inteligência artificial Overview

Bruno Duarte Corrêa – poli/usp

Thiago Dias Pastor – poli/usp

Purpose of ai Maximize fun

Emulates the reality

Not necessarily simulation

Scoped by the gameplay

Challenging

...

Desired Characteristics Be smart

Does not looks like dumb, but purposely flawed

Not intended weakness

Real time performance

Configurable

Transparent

Basic needs of ai Navigation

Perceptions

Understand / Capture a snapshot of the world

Acting

Planning

Learning / adapt

Navigation Path finding

Waypoints

Navmesh

Astar / Dijkstra

Steering behaviors

Field based navigation

Waypoint • World Discretization • Manual Waypoints • Automatic

Waypoints • Static Environment • Connect by Raytrace • Different types

NavMesh • Mesh processing algorithm • Pre-processing stage • Optimized waypoints

placement. • Sloped based • Easier to handle

• State of art

NavMesh - Recast • OpenSource Api for NavMesh Generation (c++) • Used in many commercial games • Fully integrated with Detour (Navigation System) • Recast is state of the art navigation mesh construction toolset for games. • It is automatic, which means that you can throw any level geometry at it

and you will get robust mesh out • It is fast which means swift turnaround times for level designers • Extensible and easy to integrate

Source http://code.google.com/p/recastnavigation/

Information http://www.critterai.org/book/export/html/2

PathFinding - Astar

• Find paths in the • waypoint graph • Robot Paths • Smooth filters • Bezier

• Highly Parallelizable • Can be implemented on

GPU

PathFinding - Astar F = G + H

where G = the movement cost to move

from the starting point A to a given square on the grid,

following the path generated to get there.

H = the estimated movement cost to move from that given

square on the grid to the final destination, point B.

PathFinding Summary of the A* Method 1.Add the starting square (or node) to the open list. 2.Repeat the following:

1. Look for the lowest F cost square on the open list. We refer to this as the current square.

2. Switch it to the closed list. 3. For each of the 8 squares adjacent to this current square … 4. If it is not walkable or if it is on the closed list, ignore it. Otherwise do the following. If it

isn’t on the open list, add it to the open list. Make the current square the parent of this square. Record the F, G, and H costs of the square.

5. If it is on the open list already, check to see if this path to that square is better, using G cost as the measure. If so, change the parent of the square to the current square, and recalculate the G and F scores of the square. If you are keeping your open list sorted by F score.

3.Stop when you: 1. Add the target square to the closed list, in which case the path has been found (see note

below), or 2. Fail to find the target square, and the open list is empty. In this case, there is no path.

4.Save the path. Working backwards from the target square, go from each square to its parent square until you reach the starting square. That is your path.

PathFinding

Steering Behaviors • Flow Driven • Atraction and Repulsion • No need for environment

discretization • Dynamic calculation • More human like (natural)

walking style • Characters can get stuck !!! • Swarm Behavior !!!

Basic Behaviors Seek and Flee

Pursuit and Evasion

Basic Behaviors Collision avoidance Path Pursuit

Follow the Leader Advanced Collisions Avoidance

Steering Behaviors

• Combine all results like a Resultant Force • Use physic Integration to get: • Force • Aceleration • Velocity • Space

• Can be integrated with common physics APIs

Basic needs of ai Navigation

Perceptions

Understand / Capture a snapshot of the world

Acting

Planning

Learning / adapt

Perceptions “See the world as like a human does”

Emulate human senses Response time

Sensors Sight

Field of view/Raycast Sounds Smells ?!?!

Receive Messages

Basic needs of ai Navigation

Perceptions

Understand / Capture a snapshot of the world

Acting

Planning

Learning / adapt

Finite state machine

algorithm • Simple theoretical

construct – Set of states (S) – Input vocabulary (I) – Transitional function

T(s,i)

• A way of denoting how an object can change its state over time. State: Set of properties of the environment at a time

Finite state machine

• States – E: enemy in sight – S: hear a sound – D: dead

• Events – E: see an enemy – S: hear a sound – D: die

• Action performed – On each transition – On each update in

some states (e.g. attack)

Non deterministic

Finite state machine

FSM - Limitations

Hard to Expand

Hard to Maintain

Not Flexible enough

As larger it gets as f... you are.

Behavior tree Replaces fsm

Encapsulates Actions / Logic / State

Direct acyclic graph that is traversed

Non-leaf = Condition

Leafs = Actions

Divide to conquer approach

Crysis Soldier

• Real Time Extensible • SIMPLE • Scriptable

Generic Behavior Tree

Example

Example

Example

Example

Example

Example

Example

Example

Example

END OF FRAME

Basic needs of ai Navigation

Perceptions

Understand / Capture a snapshot of the world

Acting

Planning

Learning / adapt

Planning

Goal oriented planning (GOAP)

HTN planning

STRIPS

Planning is a formalized process of searching for sequence of actions to

satisfy a goal.

GOAP • Composed by:

• Collection of states • Pre-condictions • Pos-condictions • Starting State • Goal State

• Answer the What question

• Self adaptable

GOAP-Example

GOAP-A*

Basic needs of ai Navigation

Perceptions

Understand / Capture a snapshot of the world

Acting

Planning

Learning / adapt

Supervised Learning

• Learning an input-output relationship from examples

• Tasks: regression, classification, ranking • Applications: skill estimation, behavioural cloning

Reinforcement Learning

• Learning policies from state-action-reward sequences

• Tasks: control, value estimation, policy learning • Applications: learning to drive, learning to walk, learning to fight

Unsupervised Learning

• Learning the underlying structure from examples

• Tasks: clustering, manifold learning, density estimation • Applications: modelling motion capture data, user behaviour

Learning

Neural Network • Information processing paradigm • Is inspired by the way biological nervous systems • Composed of a large number of highly

interconnected processing elements • ANNs, like people, learn by example.

Neural Network-Neuron

Neural Network

Neural Network Learning Actual algorithm for a 3-layer network (only one hidden layer): Initialize the weights in the network (often randomly)

Do

For each example e in the training set O = neural-net-output(network, e) ; forward pass

T = teacher output for e

Calculate error (T - O) at the output units Compute delta_wh for all weights from hidden layer to output layer ; backward

pass Compute delta_wi for all weights from input layer to hidden layer ; backward

pass continued

Update the weights in the network

Until all examples classified correctly or stopping criterion satisfied

Return the network

Genetic Algorithm • Genetic Algorithms are a way of solving problems

by mimicking the same processes mother nature uses.

• Use the same combination of selection, recombination and mutation to evolve a solution to a problem

• Usefull to optimize uncertain domains

Genetic Algorithm

Genetic Algorithm-Problems

• Takes too many time to converge • Hard to define the genes • Hard to define the objective function • Hard to delivery in games

Case Study – L4D

Left 4 Dead is a replayable, cooperative,

survival-horror game where four Survivors cooperate to escape

environments swarming with murderously enraged

“Infected” (ie: zombies)

Case Study – L4D

• Deliver Robust Behavior Performances

• Provide Competent Human

Player Proxies

• Promote Replayability • Generate Dramatic Game

Pacing

Deliver Robust Behavior Performances • Reactive Path Following Move towards “look ahead” point farther

down path • Use local obstacle avoidance • Good

• (Re)pathing is cheap • Local avoidance handles

small physics props, other bots, corners, etc

• Superposes well with mob flocking behavior

• Resultant motion is fluid • Bad

• Can avoid off path too much, requiring repath

• Locomotion Owns moving the actor to a new position in the environment (collision resolution, etc)

• Body •Owns animation state

• Vision •Owns line-of-sight, field of view, and “am I able to see <X>” queries. •Maintains recognized set of entities.

• Intention •Behavior Tree based

Deliver Robust Behavior Performances

• Believability/Fairness • Imperfect knowledge (senses

simulation) • Reaction times

• Trust • Survivor Bots Helps Human • Undesirable events

CANNOT occurs (Cheating) • Bots that are far from the

battle are magicaly teleported

• Survivor Bots cannot deal friendly fire

• …. • Human Behavior

Provide Competent Human Player Proxies

Promote Replayability • Complex Procedural

Population • Game session is viewed as a

skill challenge instead of a • memorization exercise • Structured Unpredictability

Structured unpredictability NavMesh

Flow Distance

Active Area Set Potential Visible Area

Generate Dramatic Game Pacing

Adaptive Dramatic Pacing algorithm • Creates peaks and valleys of intensity similar to the proven pacing success of Counter-Strike • Algorithm: • Estimate the “emotional intensity” of each Survivor • Track the max intensity of all 4 Survivors • If intensity is too high, remove major threats for awhile • Otherwise, create an interesting population of threats

AI DIRECTOR Use Survivor Intensity to modulate the Infected population • Build Up • Create full threat population until Survivor Intensity crosses peak threshold • Sustain Peak • Continue full threat population for 3-5 seconds after Survivor Intensity has peaked. Ensures minimum “build up” duration. • Peak Fade • Switch to minimal threat population (“Relax period”) and monitor Survivor Intensity until it decays out of peak range. This state is needed so current combat engagement can play out without using up entire Relax period. Peak Fade won’t allow the Relax period to start until a natural break in the action occurs. • Relax • Maintain minimal threat population for 30-45 seconds, or until Survivors have traveled far enough toward the next safe room, then resume Build Up.

Implementation considerations

Multithreading

Functional

Modular

Data driven

Cache friendly

Game interface

Level of Detail

Trends Emotion !!!

Realistic Crowd simulation

Ai in contact with the player itself

Improve ai testing

Character control and animation helper

Online ais

Interactive cut scenes

Artificial intelligence

‘Far too often, AI has been a last-minute rush job, implemented in the final two or three months of development by overcaffeinated programmers with dark circles under their eyes and thousands of other high-priority tasks to complete”

- Paul Tozour, Ion Storm