Classical Planning in Deep Latent Space: Bridging the Subsymbolic-Symbolic Boundary (1705.00154v3)

Abstract: Current domain-independent, classical planners require symbolic models of the problem domain and instance as input, resulting in a knowledge acquisition bottleneck. Meanwhile, although deep learning has achieved significant success in many fields, the knowledge is encoded in a subsymbolic representation which is incompatible with symbolic systems such as planners. We propose LatPlan, an unsupervised architecture combining deep learning and classical planning. Given only an unlabeled set of image pairs showing a subset of transitions allowed in the environment (training inputs), and a pair of images representing the initial and the goal states (planning inputs), LatPlan finds a plan to the goal state in a symbolic latent space and returns a visualized plan execution. The contribution of this paper is twofold: (1) State Autoencoder, which finds a propositional state representation of the environment using a Variational Autoencoder. It generates a discrete latent vector from the images, based on which a PDDL model can be constructed and then solved by an off-the-shelf planner. (2) Action Autoencoder / Discriminator, a neural architecture which jointly finds the action symbols and the implicit action models (preconditions/effects), and provides a successor function for the implicit graph search. We evaluate LatPlan using image-based versions of 3 planning domains: 8-puzzle, Towers of Hanoi and LightsOut.

Overview of "Classical Planning in Deep Latent Space: Bridging the Subsymbolic-Symbolic Boundary"

The paper by Masataro Asai and Alex Fukunaga explores a novel integration of deep learning with symbolic reasoning in the domain of classical planning, termed \latentplanner. The work addresses a prevalent challenge in automated planning: the knowledge acquisition bottleneck, wherein planners require a human-provided symbolic model of the planning domain. The proposed approach leverages deep learning to automate this process, bridging subsymbolic data representation and symbolic reasoning through a compelling combination of Variational Autoencoders (VAEs) and Gumbel-Softmax techniques.

Contributions and Approach

The primary contribution is the \latentplanner architecture, which uses deep neural networks to autonomously learn a symbolic representation from unlabeled image inputs. The system functions through three intelligently orchestrated phases.

1. State Autoencoder (SAE): This component utilizes a VAE with a Gumbel-Softmax latent layer to map images to propositional symbols. The bidirectional mapping ensures that symbolic states can be directly reconstructed into images, enabling visualization and comprehensibility of the plan outputs. The robustness of this system is enhanced with noise-handling capacities, ensuring symbolic planning is feasible even with imperfect sensor data.
2. Action Model Acquisition (AMA): Two complementary strategies, AMA$_1$ and AMA$_2$, are utilized to glean action models. AMA$_1$ directly compiles transitional data into a PDDL model but is impractical for large spaces due to its dependence on comprehensive transition data. Conversely, AMA$_2$ is a novel architecture that jointly learns action symbols and models from a reduced set of data, capable of providing a successor function—a pivotal requirement for state space search algorithms like A*. AMA$_2$ consists of an Action Autoencoder and an Action Discriminator, which together enable action symbol grounding and implicit action model learning.
3. Symbolic Planning and Execution: With AMA$_1$, traditional planners like Fast Downward can be applied. With AMA$_2$, a custom A* algorithm navigates the latent space using the implicit models developed.

Experimental Domains and Evaluations

The efficacy of \latentplanner is tested across several domains, including the MNIST 8-puzzle, Towers of Hanoi, and LightsOut, extending to variations with visual distortions. Each domain pushes the system's limits, evaluating both scalability and generalization across different types of visual inputs.

The paper reports robust experimental outcomes, with \latentplanner, especially under AMA$$_2`, achieving high success rates. The distinction between AMA$$_1$and AMA$_2$ clearly illustrates the benefits of implicit model representation when dealing with large state spaces. Moreover, the added noise tolerance shows potential for real-world applications where data can be inherently noisy.

Implications and Future Directions

The dual leverage of machine learning for model induction and symbolic reasoning for plan optimization positions \latentplanner at the forefront of intelligent planning systems. This research opens pathways for extending traditional AI planning techniques to new problem domains that were previously inaccessible due to the rigidity of symbolic inputs.

There are numerous potential advancements, such as integrating this system into more dynamic environments or expanding to domains with partially observable states. The methods proposed could also stimulate further exploration into robust PDDL model generation directly from sensory data in various robotics and autonomous agent applications. Additionally, with the increasing attention on explainability in AI, further research could focus on developing interpretable propositional symbols directly linked to real-world semantics.

Overall, \latentplanner represents a substantial stride towards automated, scalable symbolic planning, effectively merging the data-driven prowess of deep learning with the structured rigor of classical planning algorithms.