Introduction to Automated Planning and Acting
When we build an AI agent, either physical like robot, or virtual character in video games, the Agent often needs to carry out some activities to accomplish some goals.
Planning and Acting are two main components to achieve that. We also use the term “deliberation” for the interaction between these two components. Definition of deliberation from Merriam-Webster is:
The act of thinking about or discussing something and deciding carefully : the act of deliberating
Deliberation is the connection between planning and acting. The agent thinks about what to do (activities) and how to do them before actually acting, therefore the agent acts deliberately.
Planning
The act or process of making a plan. A plan is a sequence of actions to be executed to achieve goals.
Acting
Something that the Agent does, such as move, perceive, communicate, etc.
Deliberative Acting
Deliberative acting or acting deliberately refers to performing a series of actions which can be justified to achieve the objectives. This is often done by combination of planning and acting.
Block Diagrams
This block diagram shows the components in the intelligent agent. Planning and Acting engine is responsible for the cognitive function of the agent. Beside that it also communicates with other agents, including human operator for example to receive missions (or goals) that it has to achieve.
Planning and Acting engine sends the actions which are the things that AI Platform is capable of doing and receives up-to-date world model for it to reason.
AI Platform is a model of what the agent is capable of doing, if this is a physical agent, the AI Platform also needs to talk to hardware platform which is responsible to convert the commands to the real actuation.
Hardware Platform also senses the signals from the world and sends them to AI Platform to make sense of them. AI Platform uses these signals for its perception to update the World Model for Planning and Acting engine to reason.
Our focus is on Planning and Acting engine.
Planning and Acting
Planning is a function to choose actions to achieve the goals and Acting is a function to refine the chosen actions into AI Platform Primitives, reacting to events which happens in the world and monitoring the world to ensure that it is still possible to execute current plan in current state of the world and if not, recovery actions need to be performed to get the agent back on track to achieve its goals.
Descriptive and Operational Models
These two models are used for Planning and Acting engine.
Descriptive models of actions describe the states that may happen as a result of performing an action. They are used by Planning to reason about what actions should be chosen to achieve goals.
Operational models of actions describe what actions to execute in current context (observed state) and how to organize the chosen actions. Acting relies on these models.
Continual Planning and Acting
In practice, it is impossible to model the world fully and perfectly. The agent needs to continually refine the plan while acting on it.
Several factors are:
- Too many information to model There are just too many information to model, to process such a large information the Agent will need a powerful computing resource.
- Wrong information of the world In physical agents where sensors are used to know the state of the world, the information may be inaccurate, which may be caused by the sensors inaccurate reading.
- Partial information of the world The agent can only know the world from where it is (if there are no means to communicate with other agents) and so the information that it has is partial.
- Other agents and Dynamic World Also we need to take into account the presence of other agents and the fact that the world is dynamic.
Hierarchically Organized AI Platform
A common approach in AI Planning and Acting is to organize the components hierarchically. This is usually for task-based approach, as opposed to goal-based approach.
We can think of this approach just like how we organize our tasks. Some tasks are more abstract than others.
For example if we give a task to a robot to move all boxes to room 1, we can break down the task as follows:
The top level task is the most abstract which is broken down into two sub-tasks Check Location of Boxes and Move Boxes to Room 1.
Move Boxes to Room 1 is again broken down into sub-tasks. The process continues until we reach the lowest level. The level where the AI Platform understands.
Thanks for reading this post, to know more read the next posts.