goap-ai

G.O.A.P. AI

Goal Orientated Action Planning AI

Goap-AI a library for creating AI agents which plan their actions based on a set of overall goals. It is suitable for use in games and simulations where the agent needs to respond dynamically to a changing environment.

Installation

You can install the library into your project with:

[dependencies]
goap-ai = "0.2.0"

Alternatively, you can clone the repository and set your project to use the local copy:

git clone https://github.com/FreddyWoringham/goap.git goap
cd goap

Then build the goap binary with:

cargo build --release

After which you can run the tool:

cargo run --release config.yml

Configuration

A GOAP agent is configured with a set of Goals it is trying to achieve, and a set of Actions it can perform. At each step, the agent will select which Action to perform given the current State of the environment, and how unfulfilled the Goals are that it is trying to achieve.

State

A State object is a list of key-value pairs which represent a snapshot of the environment:

state:
  energy: 50
  health: 20
  num_apples: 2
  num_uncooked_meat: 0
  num_cooked_meat: 0

Goals

Goals are essentially target values of the State which the agent is trying to achieve:

goals:
  health:
    target: 100
    kind: GreaterThanOrEqualTo
    weight: 4
  energy:
    target: 100
    kind: GreaterThanOrEqualTo
    weight: 1

When planning its actions, an agent will try and minimise "discontentment": the total difference between the current State and the Goals it is trying to achieve.

$discontentment = \sum_{i=1}^{n} weight_i \times state_i - goal_i$

Note: This is representative of GreaterThanOrEqualTo goals, but different kinds of goals will use different formulae when calculating their discontentment with the current state.

Actions

Actions are the things an agent can do to change the State of the environment in order to achieve its Goals (minimise discontentment).

actions:
  gather:
    duration: 1
    deltas:
      energy: -5
      num_apples: 5
  hunt:
    duration: 20
    deltas:
      energy: -10
      num_uncooked_meat: 3
  cook:
    duration: 2
    deltas:
      energy: -5
      num_uncooked_meat: -1
      num_cooked_meat: 1
  eat_apple:
    duration: 1
    deltas:
      energy: 5
      health: 5
      num_apples: -1
  eat_cooked_meat:
    duration: 1
    deltas:
      energy: 20
      health: 30
      num_cooked_meat: -1
  rest:
    duration: 5
    deltas:
      energy: 10
  wait:
    duration: 1
    deltas:
      energy: -1

Planner

Algorithms

Our planner offers three primary algorithms, each optimized for specific scenarios and requirements:

• Traditional Planning
• Efficiency-Based Planning
• Hybrid Planning

Each algorithm operates in one of two solution modes:

• Fast: Prioritizes speed, delivering quick but potentially suboptimal plans.
• Best: Seeks the most optimal plan, which may require more computational resources and time.

1. Traditional Planning

Description:
Utilizes an exhaustive depth-first search approach to explore all possible action sequences up to a specified max_depth. This method aims to minimize total discontentment without considering the time efficiency of actions.

Use Case:
Ideal for scenarios where achieving the lowest possible discontentment is crucial, and there are no strict time constraints.

2. Efficiency-Based Planning

Description:
Focuses on maximizing discontentment reduction per unit of time. This approach favors actions that offer the most significant improvement in discontentment for the least time investment.

Use Case:
Perfect for time-critical tasks where actions have varying durations, and quick responsiveness is essential.

3. Hybrid Planning

Description:
Combines both traditional and efficiency-based strategies. It dynamically switches between minimizing discontentment and optimizing efficiency based on the current planning context.

Use Case:
Suitable for complex environments where both optimal discontentment reduction and time efficiency are important, allowing the planner to adapt to changing priorities.

Solution Modes

Fast

Description:
Employs heuristic-based algorithms (e.g., A*) to generate plans swiftly. While faster, these plans may not always be the most optimal.

Advantages:

• Quick response time
• Suitable for real-time applications

Trade-offs:

• Potentially suboptimal plans compared to exhaustive search methods

Best

Description:
Uses exhaustive search techniques to explore all possible action sequences up to max_depth, ensuring the most optimal plan is found.

Advantages:

• Guarantees the lowest possible discontentment

Trade-offs:

• Higher computational cost
• Longer planning time

Selecting the Right Algorithm and Solution

Configuration

You can set the maximum number of steps the agent will take to plan a sequence of actions using the following YAML configuration:

plan:
  max_depth: 10
  algorithm: Traditional
  solution: Fast

Run

With a complete configuration such as the one below:

max_depth: 10
algorithm: Traditional
solution: Fast
state:
  energy: 50
  health: 20
  num_apples: 2
  num_uncooked_meat: 0
  num_cooked_meat: 0
goals:
  health:
    target: 100
    kind: GreaterThanOrEqualTo
    weight: 4
  energy:
    target: 100
    kind: GreaterThanOrEqualTo
    weight: 1
actions:
  gather:
    duration: 1
    deltas:
      energy: -5
      num_apples: 5
  hunt:
    duration: 20
    deltas:
      energy: -10
      num_uncooked_meat: 3
  cook:
    duration: 2
    deltas:
      energy: -5
      num_uncooked_meat: -1
      num_cooked_meat: 1
  eat_apple:
    duration: 1
    deltas:
      energy: 5
      health: 5
      num_apples: -1
  eat_cooked_meat:
    duration: 1
    deltas:
      energy: 20
      health: 30
      num_cooked_meat: -1
  rest:
    duration: 5
    deltas:
      energy: 10
  wait:
    duration: 1
    deltas:
      energy: -1

You can then generate a plan of action:

                                                 n
                                                 u
                                      n          m
                                      u          _
                                      m          u
                                      _          n
                                      c          c
                           n          o          o
                           u          o          o
                           m          k          k
                           _          e          e
     e          h          a          d          d
     n          e          p          _          _
     e          a          p          m          m
     r          l          l          e          e
     g          t          e          a          a
     y          h          s          t          t
   50         20          2          0          0      (370.00) [init]
   40 -10     20          2          0          3 +3   (380.00) hunt
   35 -5      20          2          1 +1       2 -1   (385.00) cook
   55 +20     50 +30      2          0 -1       2      (245.00) eat_cooked_meat
   50 -5      50          2          1 +1       1 -1   (250.00) cook
   70 +20     80 +30      2          0 -1       1      (110.00) eat_cooked_meat
   65 -5      80          2          1 +1       0 -1   (115.00) cook
   85 +20    110 +30      2          0 -1       0      (15.00) eat_cooked_meat
   95 +10    110          2          0          0      (5.00) rest
  105 +10    110          2          0          0      (0.00) rest

Understanding the Output:

• State Transitions: Each line represents the state of various properties (e.g., energy, health) before and after an action.
• Action Labels: Actions taken by the agent (e.g., hunt, cook, eat_cooked_meat, rest) are displayed alongside their effects.
• Discontentment Score: The number in parentheses (e.g., (370.00)) represents the total discontentment after each action.
• Action Effects: Changes to state properties are indicated with + (increase) or - (decrease) values.

Integration

If you're integrating this ai library into a game or simulation, use the following steps:

1. Define Goals that the agent is trying to achieve.
2. Generate a observable State object representing the current environment from an agent's perspective.
3. Determine the set of Actions the agent can perform to change the State.
4. Configure the planner with the desired algorithm and solution mode.
5. Generate a plan of action based on the current State, Goals, and Actions.

It's recommended that your agent should re-plan its actions after each step it takes in the environment so that it regularly adapts to changing conditions.