3. Active Algorithms

AMLGym implements a random baseline for active learning in fully observable and deterministic environments.

!pip install amlgym

Get an input domain with predicate/operator signatures and no preconditions and effects.

from amlgym.benchmarks import get_domain_path
from amlgym.util.util import empty_domain

domain = 'blocksworld'
domain_ref_path = get_domain_path(domain)
input_domain_path = empty_domain(domain_ref_path)

Create a simulator of a fully observable and deterministic environment.

from unified_planning.shortcuts import SequentialSimulator
from unified_planning.io import PDDLReader
from amlgym.benchmarks import get_problems_path

problem_path = get_problems_path(domain, kind='learning')[0]
problem = PDDLReader().parse_problem(domain_ref_path, problem_path)

env = SequentialSimulator(problem=problem)

Instantiate a random agent and learns a model by acting in the simulated environment for a maximum number of 100 steps. At each step, the agent attempts to execute an action and observes the resulting environment state. Notice that action executions can fail.

from amlgym.algorithms import get_algorithm

learner = get_algorithm('RandomAgent', max_steps=40)
model, trajectory = learner.learn(env, input_domain_path)

Print the learned model and produced trajectory.

print("##################### Learned model #####################")
print(model)

print("################# Generated trajectory ##################")
print(trajectory)