Agents Aea

Skills

Skills are the core focus of the framework's extensibility as they implement business logic to deliver economic value for the AEA. They are self-contained capabilities that AEAs can dynamically take on board, in order to expand their effectiveness in different situations.

A skill encapsulates implementations of the three abstract base classes Handler, Behaviour, Model, and is closely related with the abstract base class Task:

• Handler: each skill has zero, one or more Handler objects, each responsible for the registered messaging protocol. Handlers implement AEAs' reactive behaviour. If the AEA understands the protocol referenced in a received Envelope, the Handler reacts appropriately to the corresponding message. Each Handler is responsible for only one protocol. A Handler is also capable of dealing with internal messages (see next section).
• Behaviour: zero, one or more Behaviours encapsulate actions which further the AEAs goal and are initiated by internals of the AEA, rather than external events. Behaviours implement AEAs' pro-activeness. The framework provides a number of abstract base classes implementing different types of behaviours (e.g. cyclic/one-shot/finite-state-machine/etc.).
• Model: zero, one or more Models that inherit from the Model class. Models encapsulate custom objects which are made accessible to any part of a skill via the SkillContext.
• Task: zero, one or more Tasks encapsulate background work internal to the AEA. Task differs from the other three in that it is not a part of skills, but Tasks are declared in or from skills if a packaging approach for AEA creation is used.

A skill can read (parts of) the state of the AEA (as summarised in the AgentContext), and propose actions to the AEA according to its specific logic. As such, more than one skill could exist per protocol, competing with each other in suggesting to the AEA the best course of actions to take. In technical terms this means skills are horizontally arranged.

For instance, an AEA who is trading goods, could subscribe to more than one skill, where each skill corresponds to a different trading strategy. The skills could then read the preference and ownership state of the AEA, and independently suggest profitable transactions.

The framework places no limits on the complexity of skills. They can implement simple (e.g. if-this-then-that) or complex (e.g. a deep learning model or reinforcement learning agent).

The framework provides one default skill, called error. Additional skills can be added as packages.

Independence of Skills

Skills are horizontally layered, that is they run independently of each other. They also cannot access each other's state.

Two skills can communicate with each other in two ways. The skill context provides access via self.context.shared_state to a key-value store which allows skills to share state. A skill can also define as a callback another skill in a message to the decision maker.

Context

The skill has a SkillContext object which is shared by all Handler, Behaviour, and Model objects. The skill context also has a link to the AgentContext. The AgentContext provides read access to AEA specific information like the public key and address of the AEA, its preferences and ownership state. It also provides access to the OutBox.

This means it is possible to, at any point, grab the context and have access to the code in other parts of the skill and the AEA.

For example, in the ErrorHandler(Handler) class, the code often grabs a reference to its context and by doing so can access initialised and running framework objects such as an OutBox for putting messages into.

self.context.outbox.put_message(message=reply)

Moreover, you can read/write to the agent context namespace by accessing the attribute SkillContext.namespace.

Importantly, however, a skill does not have access to the context of another skill or protected AEA components like the DecisionMaker.

What to Code

Each of the skill classes has three methods that must be implemented. All of them include a setup() and teardown() method which the developer must implement.

Then there is a specific method that the framework requires for each class.

handlers.py

There can be none, one or more Handler class per skill.

Handler classes can receive Message objects of one protocol type only. However, Handler classes can send Envelope objects of any type of protocol they require.

• handle(self, message: Message): is where the skill receives a Message of the specified protocol and decides what to do with it.

A handler can be registered in one way:

• By declaring it in the skill configuration file skill.yaml (see below).

It is possible to register new handlers dynamically by enqueuing new Handler instances in the queue context.new_handlers, e.g. in a skill component we can write:

self.context.new_handlers.put(MyHandler(name="my_handler", skill_context=self.context))

behaviours.py

Conceptually, a Behaviour class contains the business logic specific to initial actions initiated by the AEA rather than reactions to other events.

There can be one or more Behaviour classes per skill. The developer must create a subclass from the abstract class Behaviour to create a new Behaviour.

• act(self): is how the framework calls the Behaviour code.

A behaviour can be registered in two ways:

• By declaring it in the skill configuration file skill.yaml (see below)
• In any part of the code of the skill, by enqueuing new Behaviour instances in the queue context.new_behaviours. In that case, setupis not called by the framework, as the behaviour will be added after the AEA setup is complete.

The framework supports different types of behaviours:

• OneShotBehaviour: this behaviour is executed only once.
• TickerBehaviour: the act() method is called every tick_interval. E.g. if the TickerBehaviour subclass is instantiated

There is another category of behaviours, called CompositeBehaviour:

• SequenceBehaviour: a sequence of Behaviour classes, executed one after the other.
• FSMBehaviour: a state machine of State behaviours. A state is in charge of scheduling the next state.

If your behaviour fits one of the above, we suggest subclassing your behaviour class with that behaviour class. Otherwise, you can always subclass the general-purpose Behaviour class.

Follows an example of a custom behaviour:

from aea.skills.behaviours import OneShotBehaviour

class HelloWorldBehaviour(OneShotBehaviour):
        
    def setup(self):
        """This method is called once, when the behaviour gets loaded."""

    def act(self):
        """This methods is called in every iteration of the agent main loop."""
        print("Hello, World!")

    def teardown(self):
        """This method is called once, when the behaviour is teared down."""
    

If we want to register this behaviour dynamically, in any part of the skill code (i.e. wherever the skill context is available), we can write:

self.context.new_behaviours.put(HelloWorldBehaviour(name="hello_world", skill_context=self.context))

Or, equivalently to the previous two code blocks:

def hello():
    print("Hello, World!")

self.context.new_behaviours.put(OneShotBehaviour(act=hello, name="hello_world", skill_context=self.context))

The callable passed to the act parameter is equivalent to the implementation of the act method described above.

The framework is then in charge of registering the behaviour and scheduling it for execution.

tasks.py

Conceptually, a Task is where the developer codes any internal tasks the AEA requires.

There can be one or more Task classes per skill. The developer subclasses abstract class Task to create a new Task.

• execute(self): is how the framework calls a Task.

The Task class implements the functor pattern. An instance of the Task class can be invoked as if it were an ordinary function. Once completed, it will store the result in the property result. Raises error if the task has not been executed yet, or an error occurred during computation.

We suggest using the task_manager, accessible through the skill context, to manage long-running tasks. The task manager uses multiprocessing to schedule tasks, so be aware that the changes on the task object will not be updated.

Here's an example:

In tasks.py:

from aea.skills.tasks import Task


def nth_prime_number(n: int) -> int:
    """A naive algorithm to find the n_th prime number."""
    assert n > 0
    primes = [2]
    num = 3
    while len(primes) < n:
        for p in primes:
            if num % p == 0:
                break
        else:
            primes.append(num)
        num += 2
    return primes[-1]


class LongTask(Task):

    def setup(self):
        """Set the task up before execution."""

    def execute(self, n: int):
        return nth_prime_number(n)

    def teardown(self):
        """Clean the task up after execution."""

In behaviours.py:

from aea.skills.behaviours import TickerBehaviour
from packages.my_author_name.skills.my_skill.tasks import LongTask


class MyBehaviour(TickerBehaviour):

    def setup(self):
        """Setup behaviour."""
        my_task = LongTask()
        task_id = self.context.task_manager.enqueue_task(my_task, args=(10000, ))
        self.async_result = self.context.task_manager.get_task_result(task_id)  # type: multiprocessing.pool.AsyncResult

    def act(self):
        """Act implementation."""
        if self.async_result.ready() is False:
            print("The task is not finished yet.")
        else:
            completed_task = self.async_result.get()  # type: LongTask
            print("The result is:", completed_task.result)
            # Stop the skill
            self.context.is_active = False

    def teardown(self):
        """Teardown behaviour."""

Models

The developer might want to add other classes on the context level which are shared equally across the Handler, Behaviour and Task classes. To this end, the developer can subclass an abstract Model. These models are made available on the context level upon initialization of the AEA.

Say, the developer has a class called SomeModel

class SomeModel(Model):
    ...

Then, an instance of this class is available on the context level like so:

some_model = self.context.some_model

Skill Configuration

Each skill has a skill.yaml configuration file which lists all Behaviour, Handler, and Task objects pertaining to the skill.

It also details the protocol types used in the skill and points to shared modules, i.e. modules of type Model, which allow custom classes within the skill to be accessible in the skill context.

name: echo
authors: fetchai
version: 0.1.0
license: Apache-2.0
behaviours:
  echo:
    class_name: EchoBehaviour
    args:
      tick_interval: 1.0
handlers:
  echo:
    class_name: EchoHandler
    args:
      foo: bar
models: {}
dependencies: {}
protocols:
- fetchai/default:1.1.7

Error Skill

All AEAs have a default error skill that contains error handling code for a number of scenarios:

• Received envelopes with unsupported protocols
• Received envelopes with unsupported skills (i.e. protocols for which no handler is registered)
• Envelopes with decoding errors
• Invalid messages with respect to the registered protocol

The error skill relies on the fetchai/default:1.0.0 protocol which provides error codes for the above.

Custom Error Handler

The framework implements a default ErrorHandler. You can implement your own and mount it. The easiest way to do this is to run the following command to scaffold a custom ErrorHandler:

aea scaffold error-handler

Now you will see a file called error_handler.py in the AEA project root. You can then implement your own custom logic to process messages.

Source: fetchai/agents-aea — distributed by TomeVault.