Autonomous Systems -- An Architectural Characterization (1811.10277v1)

Abstract: The concept of autonomy is key to the IoT vision promising increasing integration of smart services and systems minimizing human intervention. This vision challenges our capability to build complex open trustworthy autonomous systems. We lack a rigorous common semantic framework for autonomous systems. It is remarkable that the debate about autonomous vehicles focuses almost exclusively on AI and learning techniques while it ignores many other equally important autonomous system design issues. Autonomous systems involve agents and objects coordinated in some common environment so that their collective behavior meets a set of global goals. We propose a general computational model combining a system architecture model and an agent model. The architecture model allows expression of dynamic reconfigurable multi-mode coordination between components. The agent model consists of five interacting modules implementing each one a characteristic function: Perception, Reflection, Goal management, Planning and Self-adaptation. It determines a concept of autonomic complexity accounting for the specific difficulty to build autonomous systems. We emphasize that the main characteristic of autonomous systems is their ability to handle knowledge and adaptively respond to environment changes. We advocate that autonomy should be associated with functionality and not with specific techniques. Machine learning is essential for autonomy although it can meet only a small portion of the needs implied by autonomous system design. We conclude that autonomy is a kind of broad intelligence. Building trustworthy and optimal autonomous systems goes far beyond the AI challenge.

The paper by Sifakis proposes a computational model for autonomous systems, emphasizing the ability to handle knowledge and adapt to environmental changes. The author argues that autonomy should be associated with functionality rather than specific techniques, with machine learning being only a part of autonomous system design. The paper addresses the lack of a common semantic framework for autonomous systems and the prevalent focus on AI (Artificial Intelligence) and learning techniques in discussions about autonomous vehicles, while neglecting other design considerations.

The author introduces a computational model that combines a system architecture model with an agent model. The agent model consists of five interacting modules: Perception, Reflection, Goal management, Planning, and Self-adaptation.

The paper uses examples such as a thermostat, an automatic train shuttle, a chess-playing robot, a soccer-playing robot, and a robocar to illustrate the concept of autonomy. These examples highlight the differences in complexity and intricacy of environments and goals, as well as the associated decision-making processes.

Agent Environment

The agent continuously interacts with its environment through sensors and actuators. The complexity and intricacy of environments and goals vary across different automated systems, influencing the decision-making process.

Agent Goals and Plan Generation

Agents behave as controllers, acting on their environment to achieve specific goals. The environment can be modeled as a state machine with controllable and uncontrollable actions. For infinite or complex environments, generating an explicit controller is not possible, and the existence of plans cannot be theoretically guaranteed. Finite-horizon plans are computed online using heuristics and precomputed plan skeletons.

Main Aspects of Autonomy

Autonomy is defined as the capacity of an agent to achieve coordinated goals independently, adapting to environmental variations. It combines perception, reflection, goal management, planning, and self-adaptation. The level of autonomy characterizes the relation between machine-empowered and human-assisted autonomy.

A Computational Model for Autonomous Systems

The paper presents an architecture model inspired by the BIP coordination language which has been studied and implemented in declarative and imperative formalisms [3, 4, 5]. The system architecture model involves agents and objects, with agents having computational capabilities to change the states of objects and coordinate to enforce global system goals.

System Architecture Model

A system model is a collection of architecture motifs. A motif is a world where dynamically changing sets of agents and objects exist and is equipped with a map represented by a graph. The position of an agent $a$ or object $o$ is given by a partial address function $@$, where $@(a)$ and $@(o)$ are the nodes of the map where $a$ and $o$ are located, respectively.

• $a$: agent
• $o$: object

The dynamics of the system are described by a transition relation between configurations. A configuration is the set of the states of its components and their corresponding addresses on the map. Configurations change when events occur as a result of agent coordination through interaction or configuration rules.

Computational Model for Agents

The agent computational model consists of four main modules and a Knowledge Repository.

• Knowledge Repository: Contains knowledge used by other modules for interpreting sensory information, building the environment model, and goal management. Knowledge can be declarative or procedural.
• Perception Module: Extracts relevant information from stimuli provided by sensors, using learning techniques or analysis and recognition processes.
• Reflection Module: Uses information from the Perception module to build/update a model of the agent's environment.
• Decision Module: Includes a Goal Manager for handling agent's goals and a Planner for generating plans that implement particular goals.
• Self-Adaptation Module: Supervises and coordinates all other modules, reassessing the coherency of exchanged information, creating new knowledge, and providing directives to the Goal manager.

Autonomous System Design Complexity Issues

The choice of autonomy level for risk-benefit optimization is determined by the required degree of trustworthiness and three types of complexity: autonomic, design, and implementation.

Autonomic Complexity

This accounts for the difficulty of building autonomous systems, influenced by factors related to perception, observability/controllability, goal management, planning, and adaptation.

Design Complexity

System design complexity characterizes the difficulty of building a system out of components and is conceptualized in terms of reactive complexity and architecture complexity. Reactive complexity characterizes the intricacy of the interaction between an agent and its environment. Agents are classified based on their reactive complexity: transformational, streaming, embedded, and cyber-physical. Architecture complexity ranges from static to self-organizing architectures.

Implementation Complexity

Implementation involves realizing the designed system model, with choices for the implementation architecture depending on how decisions are made and how information is shared. Three main types of implementation architecture are distinguished: centralized, decentralized, and distributed.

Trustworthy Autonomous Systems

The paper discusses the challenges of building trustworthy autonomous systems, given the shift from small-size centralized systems to large distributed autonomous systems. It extends the model-based approach for guaranteeing trustworthiness to autonomous systems, addressing the limitations in decomposition, formalization of high-level goals, building faithful system models, and treating machine-learning techniques as "black boxes". The paper suggests replacing individual DIR (Detection, Isolation, Recovery) mechanisms with adaptive mechanisms managing system resources globally.

The author concludes by noting the main characteristic of autonomous systems is their ability to handle knowledge and adapt to environmental changes.