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.

Articles you may be interested inNitrogen incorporation engineering and electrical properties of high-k gate dielectric ( HfO 2 and Al 2 O 3 ) films on Si (100) substrate J. Growth and characterization of hafnium silicate films prepared by UV/ozone oxidation J. Vac. Sci. Technol. A 22, 395 (2004); 10.1116/1.1649346Growth and effects of remote-plasma oxidation on thin films of HfO 2 prepared by metal-organic chemical-vapor deposition J.Characteristics of HfO 2 / HfSi x O y film as an alternative gate dielectric in metal-oxide-semiconductor devices

The initial stage of Ge heteroepitaxy on a Si(100)-2×1 surface has been investigated by low-energy electron diffraction (LEED) and Auger-electron spectroscopy (AES). The growth mode of the Ge films was studied by measuring the decrease in the Si(LVV) AES line at 92 eV with an increase in the Ge overlayer thickness. The Ge films deposited at room temperature exhibit layer-by-layer growth up to at least six monolayers. When the substrate is heated up to 350 °C, the growth mode is characterized by the Stranski–Krastanov type; i.e., the first three monolayers of growth is followed by island formation. Although these characteristics of the growth mechanism are similar to the case of Ge on Si(111)-7×7 surfaces, annealing behavior of the Ge films suggests that the bond strength between Ge and Si is stronger on Si(100) than on Si(111) surfaces. In contrast to the case of Ge on Si(111) surfaces, where the original 7×7 superstructure of the Si surface is replaced by a new 5×5 pattern at about two-monolayer coverage of Ge, the original 2×1 LEED pattern is not strongly disturbed up to about 1–2 monolayers of Ge. In addition to the detailed study on the initial stage of heteroepitaxial growth, we observed that thick Ge films deposited onto Si(100) surfaces held at 350 and 470 °C display a sharp 2×1 LEED pattern and demonstrate a single-crystal growth of a Ge(100) face on the Si(100) surface. This is further supported by a measurement of the x-ray diffraction pattern of the Ge films.

We achieved a new milestone in the difficult task of enabling agents to learn about their environment autonomously. Our neuro-symbolic architecture is trained end-to-end to produce a succinct and effective discrete state transition model from images alone. Our target representation (the Planning Domain Definition Language) is already in a form that off-the-shelf solvers can consume, and opens the door to the rich array of modern heuristic search capabilities. We demonstrate how the sophisticated innate prior we place on the learning process significantly reduces the complexity of the learned representation, and reveals a connection to the graph-theoretic notion of ``cube-like graphs'', thus opening the door to a deeper understanding of the ideal properties for learned symbolic representations. We show that the powerful domain-independent heuristics allow our system to solve visual 15-Puzzle instances which are beyond the reach of blind search, without resorting to the Reinforcement Learning approach that requires a huge amount of training on the domain-dependent reward information.

Characteristics of HfO 2 / HfSi x O y film as an alternative gate dielectric in metal-oxide-semiconductor devicesThe metal-organic ͑MO͒ chemical vapor deposition of hafnium oxide ͑HfO 2 ͒ films from a new MO precursor, Hf͑OC͑CH 3 ͒ 2 CH 2 OCH 3 ͒ 4 , was investigated. The deposition rate of HfO 2 is higher when oxygen gas is being supplied with the precursor. However, films deposited in the presence of added oxygen contain large amounts of H 2 O due to oxidation of the Hf precursor. O 2 addition process degraded HfO 2 film properties. In situ remote-plasma oxidation ͑RPO͒ is found to be effective in reducing the contaminants in HfO 2 . Leakage current in HfO 2 /Si capacitors with TiN gate electrode is also shown to be lower when deposition is without the oxygen addition and RPO treatment is subsequently performed.