<- Lecture 4 Lecture 6->

INTRODUCTION TO A.I.

Architecture 1

Lecture 5

January 24, 2007

5.1	Outline of Topics
	Coarse-grain mind model of intelligent systems
	Why do we need to study architecture in A.I.?
	Architecture and the perception-action loop
	Mutliple levels of processing
	Realtime
	Subsumption architecture: What it is
		Subsumption architecture: What it is
		Subsumption architecture Basics / Examples
		Subsumption architecture Examples II
		Subsumption Architecture: Characteristics
	Where are we headed?
		Large integrated systems
		Remote Agent
		Brain research architecture: Kosslyn & Koenig
		HONDA Asimo
	Tools

5.2	Coarse-grain mind model of intelligent systems
	Sensation	What comes in through the senses - "raw information" (processed data)
	Perception	Includes interpretation and decisions about the sensory data
	Interpretation	Interpret the perceptual data in context with knowledge and experience
	Decision	Interpretation, decision making and planning are often intermixed. A simplification includes an initial decision between interpretation and planning, the decision to act.
	Planning	The (mental and physical) act of deciding future actions
	Action	The execution of a decision/plan
	Goals	It is generally considered necesssary to model all cognitive agents as having goals; without goals comparisons between choices cannot be made and decisions become random.
	Memory	Memory is "everywhere" in a mind. Various memories serve various purposes. Nobody knows how the human memory works as a whole.

5.3	Why do we need to study architecture in A.I.?
	(sw)architecture = (hw)architecture	Architecture in A.I. serves the same purpose as architecture when building physical structures Just like in architectures for physical structures, A.I. architectures: Use blueprints for specifying materials, the main components and their interconnections Blueprints do not convey the full detail of implementation Help coordinate the construction of complex (sw) structures Specify the flow of elements (people/information) among units (rooms, buildings/ processing elements)
	A.I. systems get exceedingly complicated	A.I. is among the "new" sciences of complexity: Their subject is more complex than any studied by science to date
	intelligence = organization	Intelligence is in essence an organizational matter. An intelligence is defined by components: What the components are How they are structured How they are interconnected How they interact in realtime
	Science	Complex systems are the next challenge
	Engineering	We are headed towards larger and larger systems
	Holistic intelligence	Architecture is now among the main stumbling blocks towards understanding holistic intelligence
	All naturally intelligent systems perceive, think and act	Perception-Action is about architecture

5.4	Architecture & the Perception-Action Loop

	The old pipeline model of cognitive processing.

	The need for mental "threads"	To monitor one's own actions one has to sample the world after one has done an Act; sampling of multiple acts becomes a difficult problem in a pipeline model. To cope with this the mind must be doing "load balancing" among "multiple threads".

5.5	Multiple Levels of Processing

	The new model of layered continuous processing and execution. The top layer provides a quick response time to time-critical perceptual events (a bus coming at you); the middle layer provides a somewhat more thought-out respone set (1-3 second range) while the bottom layer provides a lot of different processing that takes quite a bit more time (e.g. remembering the name of that singer from the 50s).


	Coordination hierachies: A functional hierarchy organizes the execution of tasks according to their functions. A product hierarchy organizes production in little units, each focused on a particular product. Several types of markets exist - here two idealized versions are show, without and with brokers. De-centralized markets require more intelligence to be present in the nodes, which can be aleviated by brokers. Brokers, however, present weak points in the system: If you have a system with only 2 brokers mediating between processors and consumers/buyers, failure in these 2 points will render the system useless. Notice that in a basic program written in C++ every single character is such a potential point of failure, which is why bugs are so common in standard software. The human and animal minds are probably ... a mixture of all of these. At the gross anatomical level the brain is a functional hierarchy, with motor control and perceptual inputs in specific places (vision, for example, is always in the back of your head -- no execptions, while language is in the left hemisphere in most people).

5.6	Realtime
	Perception-Action Loop	A question of realtime

		Realtime just means "real fast". Right? Wrong. Realtime involves (at least) all of the following: 1. Responsiveness: The system’s (in this case dialog participant’s) ability to stay alert to, and respond to, incoming information. 2. Timeliness: The system’s ability to manage and meet deadlines. 3. Graceful adaptation: The system’s ability to (re)set task priorities in light of changes in resources or workload, and to rearrange tasks and replan when problems arise, e.g. in light of missed deadlines. 5. Relevance. Are the decisions relevant to the situation? 5. Speed.
	To achieve realtime	Sample the world: Monitor information sources and manage sensory apparati Load-balance threads: Monitor actions & their outcomes Compare action outcomes to goals

5.7	Subsumption architecture: What it is
	What it is	Robot control architecture system developed at the MIT AI Lab by Rodney Brooks
	why it exists	An effort to shift attention from human-level intelligence to simpler organisms, and in the process create the simplest possible architecture that could express intelligent behavior.
	how we will use it	You will study the subsumption architecture to a sufficient degree that you can implement it in Java or C++ to control the CADIA Hexapod in your final project.

5.8	Subsumption Architecture Basics / Examples
	Augmented Finite State Machines (AFSMs)	Finite State Machines, augmented with timers
	Modules (FSMs) have internal state	The internal state includes: the clock the inputs (no history) the (current) output (no history) may include "activation level"
	External environment constists of connections ("wires")	Input Inhibitor Suppressor Reset Output
	Augmented Finite State Machine (AFSM) with connections
		Suppressor: Replaces the input to the module Inhibitor: Stops the output for a given period Reset: Initialization puts the module in its original state
	Augmentation	The finite state machines are augmented with timers. The time is fixed for each I or R, per module.
	Timers	Timers enable modules to behave autonomously based on a (relative) time
	The AFSMs are arranged in "layers"	Layers separate functional parts of the architecture from each other


	Level 0 example.


	Level 0 and 1 combined.


	Level 0, 1 and 2 combined.

5.9	Subsumption Architecture Examples II

	Example subsumption architecture with 5 layers.


	Example subsumption architecture.

5.10	Subsumption Architecture: Characteristics
	Architectural characteristics	Low processing overhead Tight coupling between perception and action Modular structure Low to medium number of modules
	Behavioral characteristics	Rapid, reflexive responses Long-term and short-term goals
	Benefits	Easy to construct relatively complex systems Systems tend to work well in the real world Not CPU intensive (the architecture, that is - sensory mechanisms can still be quite expensive) Well understood framework No complicated representation of external conditions needs to be put into the creatures
	Drawbacks	Architectures tend to be uniform - it is difficult to extend them beyond controlling simple creatures. Hard to do: simultaneous (parallel) goals and complex goal structures layered planning and action control ... i.e., it is difficult to construct large architectures that have features of human-level intelligence
	Drawbacks can be overcome through use of Constructionist AI	Constructionist AI mantra: No method off-limits Use well-known mechanisms to solve problems they apply well to Expand existing architectures Develop large architectures from scratch Achive human-like features Explore integration of very different mechanisms Integrate learning into heterogeneous architectures
	Constructionist AI results in hybrid architectures	"Hybrid architectures" Typically don't buy into a particular school of thought Very often hierarchical like the subsumption architecture

5.11	Where are we headed? (next 10-20 years)
	Large integrated systems	Need to build integrated simulations of thinking systems to: Improve autonomous systems (applications) Realize proper models of intelligence (research)
	Distributed systems	No single computer is up to the task, at present
	Manual construction	Current models of cognition are hand-built Will be largely manually built for a while

5.12	Large Integrated Systems
	Increased autonomy …	…means we are moving towards systems that: Sense, recognize, classify, plan, decide, act … integrated!
	Industry examples	Increased autonomy, flexibility in factories Faster re-programmability, at a higher level
	Research examples	Remote Agent Brain research Asimo

5.13	Remote Agent
	Deep Space One	NASA satellite went into orbit on Oct. 24th 1998 Purpose: Test the value of various new technologies
	Deep Space One Remote Agent	On-board AI system: “Remote Agent” Remote Agent managed the mission autonomously


	Deep Space 1 Remote Agent architecture of main software modules.

	Remote Agent Thought Process Example	Remote Agent knows that: To turn on a power source requires “pushing a button” A “button” is a physical thing that behaves according to the laws of physics - buttons can get stuck If a physical thing gets stuck you can try to get it loose by “jiggling” it Reasonable approach to solve the problem: Jiggle the switch for the power source Remote Agent solved the problem and turned on the power source
	To think, Remote Agent uses:	Detailed model of itself, plus Heuristics, rules about the world

Kosslyn & Koenig architecture. Based on decades of brain research.

5.14	Large Complex Systems: HONDA Asimo


	Significant progress in last 5-10 years	balance, walking, running
	Asimo slated to become a household helper by 2012	Serving tea Cleaning Run errands
	Asimo is an integrated system	Perception, action, balance, planning, object recognition, etc.

Snapshots from HONDA Asimo video showing Asimo walk out of a subway station in New York City.

5.15	Tools
	The future is clear:	We need to build large-scale, distributed architectures to study and develop integrated artificial intelligence
	Significant lack of good tools to build large-scale, distributed architectures	Most tools in computer science focused on Small applications Lose networking
	Psyclone+OpenAIR	System explicitly created for large-scale AI development Builds on Constructionist AI