DECIDER-l: A System that Chooses Among Different Types of Acts

Leonard Uhr

TL;DR

This paper introduces DECIDER-1, a novel system that integrates perception, language understanding, and decision-making to choose and execute actions based on a unified input stream of objects and symbolic commands. Unlike prior systems that rigidly separate these functions, DECIDER-1 treats words and objects as a single perceptual input, enabling a more holistic approach to intelligent behavior.

What Problem Does This Paper Address?

The paper tackles a fundamental challenge in artificial intelligence: how to create systems that can not only perceive their environment and understand commands but also intelligently decide which command to obey and how to execute it, all while integrating these processes seamlessly. Current AI systems often operate in silos, with separate modules for perception, language processing, and action planning. This separation creates a disconnect, making it difficult for systems to handle the nuanced, real-world scenario where commands and environmental objects arrive through the same sensory channels and must be understood in relation to each other. The authors highlight the gap in theoretical models for how humans and even intelligent animals like chimpanzees can fluidly process mixed inputs of objects and words, infer the intent of commands, and select appropriate actions.

Methods

DECIDER-1 employs a unified approach where both objects and symbolic commands are processed through the same perceptual and linguistic mechanisms. The core of the system relies on a dynamic list called "IDEAS."

  • Unified Input Processing: DECIDER-1 takes a single "SCENE" input that contains a mixture of physical objects and verbal utterances.
  • Primitive Perceptual Transforms: The system begins by applying basic perceptual transforms to identify fundamental features (edges, gradients) and potential objects within the scene.
  • Delimiters and Decomposition: These features are used to break down the scene into tentative objects.
  • IDEAS List: This list stores potential "ideas" or transforms that the system can apply. As transforms succeed, they can imply new objects, new ideas to explore, or potential "acts" (actions).
  • Serial Application of IDEAS: The system iteratively selects the most "implied" idea from the IDEAS list and applies it. This process continues, refining the understanding of the scene and identifying potential actions.
  • Response Acts: When a transform implies a response act (e.g., Answer, Describe, Find, Move, Name) with sufficient weight, DECIDER-1 executes it. The key innovation is that these response acts are treated the same way as perceptual or linguistic transforms, allowing them to be dynamically chosen and ordered.
  • Multi-Determined Acts: The decision to perform an act is not solely based on a command but can also be influenced by perceived objects or internal needs, mirroring human decision-making.

Datasets

Not discussed in the paper. The paper focuses on the conceptual framework and behavior of DECIDER-1, using illustrative examples rather than formal training or evaluation on specific datasets.

Key Insights and Ideas

  • Unified Perception and Language: The most significant insight is that treating words and objects as a single, unified input stream is crucial for building more intelligent systems. This breaks down the artificial barrier between symbolic reasoning and sensory perception.
  • Dynamic Decision-Making: DECIDER-1's ability to dynamically choose among different types of acts (answering, describing, finding, moving, naming) based on the "IDEAS" list is a key contribution. Acts are not hardwired but emerge from the system's understanding of the input.
  • Implicit Action Inference: The system can infer actions not just from explicit commands but also from the presence of objects or internal states, demonstrating a more flexible and human-like decision process.
  • The "IDEAS" List as a Central Hub: This list acts as a powerful mechanism for managing uncertainty, exploring possibilities, and guiding the system's focus, integrating perception, understanding, and action selection.

Key Results

The paper does not present quantitative results in terms of accuracy or performance metrics on specific tasks. Instead, it demonstrates the conceptual capabilities of DECIDER-1 through qualitative examples:

  • Handling Mixed Inputs: DECIDER-1 can process scenes containing both objects (e.g., BANANA, BOX) and symbolic commands (e.g., "SARAH PUT BANANA BOX").
  • Executing Diverse Acts: The system is designed to perform various acts, including:
  • Answering: Providing stored information.
  • Describing: Naming recognized objects in the scene.
  • Finding: Locating and bracketing a specified object.
  • Moving: Manipulating objects within the scene.
  • Naming: Identifying and outputting the name of an object.
  • Example Scenarios: The paper illustrates DECIDER-1's behavior with inputs like "DESCRIBE THIS SCENE" and "FIND BOX," showing how it can correctly interpret the command and generate an appropriate output.

Limitations and Future Work

  • Simplification of Scenes: The current implementation reduces scenes to one-dimensional strings for simplicity, though the authors note that extensions to two- and three-dimensional scenes are possible.
  • Need for Primitive Transforms: The system's capabilities are dependent on the programmer providing the necessary primitive perceptual transforms, verbal rewrite rules, and action transforms.
  • Future Work: The paper mentions that DECIDER-1 is being extended to handle scenes that evolve over time, where perceiving, responding, and problem-solving will occur in parallel. This future work aims to create a more integrated and continuous intelligent system.

Summary updated

BibTeX
@inproceedings{uhr1973decider,
  author = {Leonard Uhr},
  title = {{DECIDER-l: A System that Chooses Among Different Types of Acts}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1973},
}
Figure