Goal
Oriented Action Planning
Introduction
GOAP, or goal oriented action planning is
a decision-making architecture for agents in game that allows them to plan a
sequence of actions to satisfy a particular goal. It is based on a modification
of the STRIPS language which is defined using goals and operators. It was
originally implemented by Jeff Orkin for the game F.E.A.R. GOAP is known for
the emergent behaviour that spawns from its simple components.
Why
GOAP?
Most of the games use finite state machines
for implementing AI. With complex AI, the states for FSM agents can become
really complex which eventually get to a point where adding new behaviours to
the AI is almost impossible and have too many side effects and gaps in the AI.
The difficulty in managing the complexity of FSM’s was cited as the primary
reason as to why GOAP was chosen for F.E.A.R.
GOAP can turn something like this:
Into this:
In GOAP, the actions are not connected to
each other. This way, all the actions can be focussed individually making the
code modular and easier to maintain.
Important
Terms
To learn how GOAP works, first we need to
define some important terms.
Ø States
States are used to
define a single attribute about the game world surrounding the agent, and the
attributes about the agent itself. The states are simply flags representing
some information about the world or agent. For example, an agent can have
states like LOW AMMO or HAS TARGET.
Ø Worldstates
Worldstates are
collection of flags, where each flag represents a state. There are two
different use cases for worldstates. First, it defines the current state a NPC
is in. Secondly, the user can use it to define a target worldstate. So, the
goals for the agents can be expressed as desired worldstates.
Ø Actions
An agent has to
build a sequence of actions to reach its desired worldstate or its goal from
its current worldstate. Actions have a precondition and post condition. Both
the conditions are defined as worldstates. For example, to perform an action
SHOOT, the agent current worldstate should define that the agent has a gun,
ammo and a target. The post worldstate would result into the target taking
damage. All the actions have a cost to them which are calculated dynamically.
Ø Goals
An agent has a
list of goals which, as mentioned before, are defined as world states. The
agent picks the most relevant goal dynamically from the list of goals according
to its current world state, and then basically builds a plan to fulfill that
goal.
GOAP
Planner
The GOAP planner is the code that starts
from the player current worldstate and tries to build a plan to the goal
(desired worldstate). To do that, the planner goes through the actions, checks
if the precondition for action is met and post condition is making the agent
reach closer to its desired goa while keeping cost in mind. An A* Engine is the
best way to implement the planner. A generic version of A* can be used for both
pathing and planning for GOAP.
Code
Implementation
To implement GOAP into the game, you would
need to have these important classes or structure.
Ø WorldState
structure – use to define current and target worldstates.
Ø Action
Class – An abstract class to define different actions.
Ø Goal
Class – Abstract class for all Goals.
Ø Goal
Manager – Class to update goals, goal relevancies. Calls the planner to build a plan.
Ø Goal
Planner – Use the A* engine to build a plan to achieve most relevant goal.
Benefits
of GOAP
With GOAP, large series of actions can be
created without worrying about connecting them. Actions can easily be added and
removed. The dynamic AI from GOAP behaves smarter than normal FSM. For complex
character behaviour, GOAP is a great solution for having scalable,
maintainable, and reusable decision making systems.
