Intelligence without Reason Rodney A. Brooks CS790X Anil Shankar
Overview of the talk Status-check on research in AI Intelligence without explicit reasoning systems Influence of various disciplines and technology on the development of AI Situatedness, Embodiment, Intelligence and Emergence CS790X Anil Shankar
Robotics Static environments Off board computation Sense-Model-Plan-Act architectures (SMPA) Assuming that the static world can scale to the real dynamic world Were these robots “intelligent”? CS790X Anil Shankar
Re-think Intelligence Do we always problem-solve and plan? An agent’s internal representation compared with real-world object representation Where should the agents be? Can an agent have goals and beliefs? So how do we re-think then ? CS790X Anil Shankar
The new manifesto Situatedness (S) Embodiment (E) Intelligence (I) Emergence (E) Compare SEIE with SMPA Check your computer for intelligence CS790X Anil Shankar
Us and Them Silicon based machines Biological machines Von Neumann architecture Biological machines Low speed, massively parallel, fixed and bounded network topology, redundancies in design What would the classical AI guys say? CS790X Anil Shankar
Classical A.I Turing Test Chess Dartmouth Conference AI techniques Allowed disembodiment Chess What about Go? Dartmouth Conference Search AI techniques Search, Pattern recognition, learning, planning and induction (disembodied and non-situated, reliance on performance increases Where did all these ideas come from ? CS790X Anil Shankar
Other Disciplines Cybernetics Abstraction Knowledge Representation Organism and it’s environment should be modeled together (situatedness) Abstraction Blocks world, controlled environments, Shakey, internal models, complacence with performance in static environments Knowledge Representation Represent knowledge, problem-solve, learn …ungrounded! CS790X Anil Shankar
Other disciplines (2) Vision Parallelism Biology Reconstruct static external world as a three dimensional model Parallelism Neural networks, no situatedness, hand-crafted problems, real-world performance missing Biology Use ethology to make an ungrounded assumption about hierarchical models of thinking/intelligence CS790X Anil Shankar
Other disciplines (3) Psychology Marr’s view of vision maybe different from biological vision Representation of knowledge as Central storage (concepts, individuals, categories, goals, intentions, etc.) Knowledge stored independent of the circumstances in which it is acquired Modality-specific organization of meaning CS790X Anil Shankar
Other disciplines (4) Neuroscience What about the hormones? Do we know enough about the neurological organization simple creatures? Do we want to consider something that might actually work? CS790X Anil Shankar
Brave New World Situatedness Embodiment Intelligence Emergence The world is its own best model Embodiment The world grounds regress Intelligence Intelligence is determined by the dynamics of interaction with the world Emergence Intelligence is in the eye of the observer Will these work ? CS790X Anil Shankar
Brooks’ Approach Situatedness Embodiment Highly reactive architectures with manipulable representations No symbols and decentralized computation What do we need next? CS790X Anil Shankar
Domain Principles Complete integrated intelligent autonomous agents Embodiment in the real world Efficient performance in dynamic environments Operate on time-scales in proportion to that used by humans How do we realize them ? CS790X Anil Shankar
Computation Principles Asynchronous network having active computational components No implicit semantics in exchanged messages Asynchronously connected sensors and actuators to two-sided buffers What will these ideas help us realize? CS790X Anil Shankar
Some consequences A state enabled system and not just a reactive one Bounded search space Simple data structures No implicit separation of data and computation Practice and Principles ? CS790X Anil Shankar
More on Brooks’ robots No central model, no central control locus Network components can perform more than one function Behavior specific networks, build and test method No hierarchical arrangement, parallel operation of behaviors (layers) Use the world itself as a communication medium Simpler design, on-board computation, miniaturization possible Limitations Power, computational capability The real robots please CS790X Anil Shankar
A few specific robots Allen Herbert Toto Reactive, sonar, non-reactive goal selecting layer, same computational mechanism for both reactive and non-reactive components Herbert World as it’s own model, opportunistic control system, adapt to dynamic changes Toto Extract only relevant representations, decentralized, active-maps Complex goal-directed and intentional behavior with no long term internal state Everything is not peachy CS790X Anil Shankar
A few issues Complexity Learning Behaviors Environment, sensors and actuators, layers Learning Representations for a task, calibration, interaction of modules, new modules Behaviors Specification, number, interaction What else is there to do next? CS790X Anil Shankar
Issues Convergence Synthesis Complexity Learning Coherence Relevance Adequacy Representation Emergence Communication Cooperation Interference Density Individuality Almost done CS790X Anil Shankar
Status-check on research in AI Main Points Status-check on research in AI Intelligence without explicit reasoning systems, emergent property and evolutionary basis Influence of various disciplines and technology on the development of AI Situatedness, Embodiment, Intelligence and Emergence Questions ? Comments? Suggestions ? CS790X Anil Shankar