AgentSpeak - Introduction

In the lecture, you learned about the BDI agent architecture. In this seminar, we will talk about AgentSpeak and Jason. AgentSpeak is a programming language for implementing agents in the BDI architecture. Jason is its interpreter. Additionally, Jason also provides simple definitions of environment, where the agents can move and therefore it is possible to create rather simply even more complicated simulations.

AgentSpeak is an interpreted language with a Prolog-like syntax and functionality. Its interpreter is called Jason and can be downloaded from http://jason.sourceforge.net/wp/. Installation is simple - just unzip the archive to any directory you like. In order to work with Jason you need an IDE. The distribution of Jason contains a version of jEdit, which supports the AgentSpeak and Jason projects. There is also an Eclipse plugin, which adds Jason support, but it seems to be mostly abandoned. The installation instructions for the plugin are available at http://jason.sourceforge.net/mini-tutorial/eclipse-plugin/, however, I recommend using jEdit.

Cleaning Robots Example

The Jason interpreter is distributed with a number of examples (in the examples directory). We will show the basics using the cleaning-robots example.

This example contains two robots. One of them moves through the environment and collects garbage, the other one sits in the middle and destroys all the garbage the former robots takes to it (more precisely, all garbage at its location). The environment is represented by a 7x7 grid and provides a few basic actions to the agents.

• next(slot) – moves the agent to the next slot
• moveTowards(x,y) – moves the agent one slot closer to the given coordinates
• pick(garb) – picks the garbage (can fail)
• drop(garb) – drops the garbage
• burn(garb) – burn the garbage at the r2 location (that’s the agent which burns the garbage)

The environment also provides the following information

• pos(R,X,Y) – robot R is at position (X,Y)
• garbage(R) – there is garbage at the position of robot R

Open the example in jEdit - the main file for the example is called mars.mas2j and it describes the whole multi-agent system. The basic syntax is of .mas2j files is rather simple, you should not have any problems understanding it - it just specifies the name of the system, which infrastructure to use for communication (we will always use Centralised, but Jason also supports jade), the main environment class, and a list of agents. Later, we will also see that we can specify parameters to the agents and other things.

In order to run the examples in jEdit you need to set path to the jason.jar file and to ant-libs. You can find these settings under Plugins -> Plugin Options -> Jason. Finally, you need to click the green arrow in the bottom right corner.

Environment Implementation

From an implementation point of view, the environment for the agent is created as an implementation of the Environment class. It is important to implement the methods init(String []) and executeAction(String ag, Structure action). The environment typically uses a model, in this case, a GridWorldModel which can be used for environments implemented as grids. In the executeAction method, you can see how we can implement the actions in the environment and how to parse the parameters of such actions. The method also calls the updatePercepts method that updates the percepts of the agents.

Jason can also work with other frameworks for the implementation of more complicated environments, such as Cartago. However, we will not need these frameworks and will not discuss them any deeper.

Agent Implementation

Each AgentSpeak program consists of several parts – rules, initial beliefs, and plans and goals.

Initial Beliefs and Rules

Rules in AgentSpeak express relationship between beliefs, or any other general information about the world. For example, if we want to write a rule which holds, if the robot r1 is at the same location as robot P (and the locations of robots are represented as pos(R, X, Y) in beliefs) you can define a predicate at(P) as

at(P) :- pos(P,X,Y) & pos(r1,X,Y).

Note that the syntax of the rules is similar to the syntax of Prolog. The only difference is that instead of comma the atoms are delimited by &. The AgentSpeak interpreter even makes unifications when applying a rule and the reasoning is also similar to Prolog.

Both rules and atoms provided in the beginning form the set of initial beliefs, which can later change during the life of the agent. The agent can act and reason based on these rules and beliefs.

Beliefs are automatically updated when the agents learns a new information from the environment or from another agent. Each belief can also have an annotation (written in square brackets) which describes the source of the belief, e.g. likes(peter, music)[source(martin)], denotes that the agent beliefs peter likes music, because the agent martin told him so. Annotations can be unified in the same way as beliefs.

Plans

The behavior of an agent is given mainly by its plans. The description of plan in Jason is again very simple. Each plan consists of three parts – trigger, context and the body.

The trigger tells when the plan should start. Syntactically, it is similar to the head of a predicate in Prolog with a few symbols added in front of the head. These symbols express whether the plan should execute after the addition (+) or removal (-) of a belief (nothing) or goal (!, or ? for query goals). In the example agent, you can find a plan which is executed if the garbage(r1) belief is added. Its trigger is +garbage(r1). You can also find a trigger for the addition of the initial plan check(slots), which is +!check(slots). Again, unification is used with triggers in the same way as with the rules. The unified variables are bound in both the context and the body of the plan.

Next part of plan is a context (divided by : from the trigger). It is in fact a condition which must hold before the plan can be executed. The syntax is similar to the syntax of the body of a rule. Again, in the sample agent, you can find the rule

+!ensure_pick(S) : garbage(r1)
[...]

The part after the colon is the context.

There are also contexts, which contain the keyword not. It means, that the agent does not believe the condition is true. There is also the ~ operator, which means that the agent believes the condition is not true.

The last part of each plan is its body. It is delimited by <- from the context and contains a list of actions the agent should perform. There are several kinds of actions. The actions provided by the environment are written without any qualifier (e.g. pick(garb)), on the other hand, internal actions of the agents start with a dot (.print(“Hello world”)).

There are special types of actions to add a new plan (they start with ! or ? for query plans), add (+) or remove (-) a belief. We can also use (-+) as a shorthand for the update of a belief. If you want to ask the belief base, you can use the ? operator (it is in fact addition of a query goal). The query goal is a special type of plan. It is handy, if we need to find out something, which is not in the belief base (e.g. an agent sitting in the living room wants to know, whether there is beer in the fridge, it needs to move to the fridge first).

Tasks

The main goal of this lesson is for you to understand the basics of AgentSpeak, so, follow the text above, download Jason and check the source codes of the cleaning-agents example. Afterwards, you can try and modify the sources in some of the following ways:

1. Add another burner robot to the environment.
2. Update the cleaning robot so that it carries to garbage to the closest burner.
3. Add more cleaning agents. 

There are more interesting simple examples, another one is called blocks-world where the goal is for a single agent to build towers from blocks. The source code of the agents is nicely commented, so I recommend checking it out.