MAcq: The Model Acquisition Toolkit

This library is a collection of tools for planning-like action model acquisition from state trace data. It contains a reimplementation from many existing works, and generalizes some of them to new settings.

Example Usage

from macq import generate, extract
from macq.observation import IdentityObservation, AtomicPartialObservation

# get a domain-specific generator: uses api.planning.domains problem_id/
# generate 100 traces of length 20 using vanilla sampling
traces = generate.pddl.VanillaSampling(problem_id = 123, plan_len = 20, num_traces = 100).traces

traces.generate_more(10)

action = traces[0][0].action
traces.get_usage(action)
# [0.05, 0.05, ..., 0.05]

trace = traces[0]
len(trace)
# 20

trace.fluents
trace.actions
trace.get_pre_states(action) # get the state before each occurance of action
trace.get_post_states(action) # state after each occurance of action
trace.get_total_cost()

Survey

You can find the full scope of papers considered in the survey (implemented and otherwise) at http://macq.planning.domains . This repository of model acquisition techniques will be added to over time.

Related Survey Papers

• A Review of Machine Learning for Automated Planning (see Fig 2)
• A Review of Learning Planning Action Models (see Tbl 3)

Citing this work

@inproceedings{macq-keps-2022,
  author    = {Ethan Callanan and Rebecca De Venezia and Victoria Armstrong and Alison Paredes and Tathagata Chakraborti and Christian Muise},
  title     = {MACQ: A Holistic View of Model Acquisition Techniques},
  booktitle = {The ICAPS Workshop on Knowledge Engineering for Planning and Scheduling (KEPS)},
  year      = {2022}
}

General API

Trace Generation

These are the various methods implemented to generate the base trace data.

Tokenization

Once trace data is loaded, you can process the traces to produce lists of observations. The methods range from the identity observation (constaining the same data as original traces) to noisy and/or partially observable observations.

Extraction Techniques

Depending on the observation type, different extraction techniques can be used to extract the relevant information from the observations. These are currently the techniques implemented: