Multi-Level Organization for Learning A Multi-Level Organization of Semantic Primitives for Learning Models of Environment Autonomously from Continuous Data for Design

In this paper, we present an overview of the project on Autonomous Learning Design Agents which began recently. An agent that addresses the task of Design by Autonomous Learning (DAL) builds an abstract model of the environment from the sensory data with the goal to modify the environment to meet a new set of functionalities. Any agent that addresses the DAL task is confronted with a task of transforming the underlying causes that produce the continuous data of the environment to produce another set of continuous data. In our work, we explore the possibility of such a transformation through a hierarchical order of discretizations each of which allows the agent to act differently, and yet allows the agent to draw some global conclusions. The machine learning research has identified formulation of environment models through multiple levels of abstractions (Pierce and Kuipers 1997). Such a formulation brings to focus some issues that need to be addressed for DAL task. 1. How does an agent recognize satisfaction of a global and abstract goal from continuous data? 2. How can an agent maintain the grounding of discretized model at different levels, in continuous data? 3. How can an agent decide upon local actions based upon global models? 4. How can an agent generate discretized models and yet open to new continuous data? 5. How does an agent decide upon environmental transformations that result in new continuous data? In a theory driven hybrid system, the relationship between the discrete model and the continuous data is supplied by the theory (Zhao 1997). The data driven approaches are able to cope with continuous data in a large number of situations. They are limited to consider abstract global goals at different levels which is often required in design (Chapman 1991). Our approach integrates the strengths of a data-driven and theory-driven approaches and has the major contributing points (Prabhakar 1999). 1. The models, at each level, are built by incorporating the interaction mechanisms of the agent into the continuous data. The interaction mechanisms allow the discretized data to be grounded in continuous data. These models are predictive. 2. Due to such an incorporation, the models can incorporate new continuous data into them. 3. The models are based on the composition of data elements, rather than the concepts of the agent. This adds flexibility to the model to building. 4. The abstraction processes, at different levels, transform the continuous …