LECTURE 5: These problems have led some researchers to question the - PDF document

Reactive Architectures � There are many unsolved (some would say insoluble) problems associated with symbolic AI LECTURE 5: � These problems have led some researchers to question the viability of the whole paradigm, and to REACTIVE AND HYBRID the development of reactive architectures � Although united by a belief that the assumptions ARCHITECTURES underpinning mainstream AI are in some sense An Introduction to MultiAgent Systems wrong, reactive agent researchers use many http://www.csc.liv.ac.uk/~mjw/pubs/imas different techniques � In this presentation, we start by reviewing the work of one of the most vocal critics of mainstream AI: Rodney Brooks 5-1 5-2 Brooks – behavior languages Brooks – behavior languages He identifies two key ideas that have � Brooks has put forward three theses: informed his research: � Intelligent behavior can be generated without Situatedness and embodiment: ‘Real’ 1. 1. explicit representations of the kind that symbolic intelligence is situated in the world, not in AI proposes disembodied systems such as theorem provers or expert systems Intelligent behavior can be generated without 2. explicit abstract reasoning of the kind that Intelligence and emergence: ‘Intelligent’ behavior 2. symbolic AI proposes arises as a result of an agent’s interaction with its environment. Also, intelligence is ‘in the eye of Intelligence is an emergent property of certain 3. the beholder’; it is not an innate, isolated complex systems property 5-3 5-4 A Traditional Decomposition of a Mobile Brooks – behavior languages � To illustrate his ideas, Brooks built some based on Robot Control System into Functional his subsumption architecture Modules � A subsumption architecture is a hierarchy of task- accomplishing behaviors � Each behavior is a rather simple rule-like structure � Each behavior ‘competes’ with others to exercise control over the agent � Lower layers represent more primitive kinds of behavior (such as avoiding obstacles), and have precedence over layers further up the hierarchy � The resulting systems are, in terms of the amount of computation they do, extremely simple � Some of the robots do tasks that would be impressive if they were accomplished by symbolic From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 AI systems 5-5 5-6 1

A Decomposition of a Mobile Robot Layered Control in the Subsumption Control System Based on Task Achieving Architecture Behaviors From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 5-7 5-8 Example of a Module – Avoid Schematic of a Module From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 5-9 5-10 Levels 0, 1, and 2 Control Systems Steels’ Mars Explorer � Steels’ Mars explorer system, using the subsumption architecture, achieves near- optimal cooperative performance in simulated ‘rock gathering on Mars’ domain: The objective is to explore a distant planet, and in particular, to collect sample of a precious rock. The location of the samples is not known in advance, but it is known that they tend to be clustered. From Brooks, “A Robust Layered Control System for a Mobile Robot”, 1985 5-11 5-12 2

Steels’ Mars Explorer Rules Steels’ Mars Explorer Rules � For individual (non-cooperative) agents, the lowest-level behavior, (and hence the behavior with the highest � Agents will collect samples they find: “priority”) is obstacle avoidance: if detect a sample then pick sample up (4) if detect an obstacle then change direction (1) � An agent with “nothing better to do” will explore � Any samples carried by agents are dropped back at the randomly: mother-ship: if true then move randomly (5) if carrying samples and at the base then drop samples (2) � Agents carrying samples will return to the mother-ship: if carrying samples and not at the base then travel up gradient (3) 5-13 5-14 Situated Automata Situated Automata � A sophisticated approach is that of Rosenschein � The logic used to specify an agent is and Kaelbling essentially a modal logic of knowledge � In their situated automata paradigm, an agent is � The technique depends upon the possibility specified in a rule-like (declarative) language, and of giving the worlds in possible worlds this specification is then compiled down to a digital semantics a concrete interpretation in terms machine, which satisfies the declarative of the states of an automaton specification � “[An agent]… x is said to carry the information � This digital machine can operate in a provable that P in world state s , written s╞ K(x,P) , if for time bound all world states in which x has the same value � Reasoning is done off line , at compile time , rather as it does in s , the proposition P is true.” than online at run time [Kaelbling and Rosenschein, 1990] 5-15 5-16 Situated Automata Circuit Model of a Finite-State Machine � An agent is specified in terms of two components: perception and action � Two programs are then used to synthesize agents � RULER is used to specify the perception component of an agent � GAPPS is used to specify the action component f = state update function s = internal state g = output function From Rosenschein and Kaelbling, “A Situated View of Representation and Control”, 1994 5-17 5-18 3

RULER – Situated Automata GAPPS – Situated Automata � The GAPPS program takes as its input � RULER takes as its input three components � A set of goal reduction rules , (essentially rules that � “[A] specification of the semantics of the [agent's] encode information about how goals can be inputs (‘whenever bit 1 is on, it is raining’); a set of achieved), and static facts (‘whenever it is raining, the ground is � a top level goal wet’); and a specification of the state transitions of the world (‘if the ground is wet, it stays wet until the � Then it generates a program that can be sun comes out’). The programmer then specifies the translated into a digital circuit in order to desired semantics for the output (‘if this bit is on, the realize the goal ground is wet’), and the compiler ... [synthesizes] a � The generated circuit does not represent or circuit whose output will have the correct semantics. manipulate symbolic expressions; all symbolic ... All that declarative ‘knowledge’ has been reduced manipulation is done at compile time to a very simple circuit.” [Kaelbling, 1991] 5-19 5-20 Circuit Model of a Finite-State Machine Situated Automata � The theoretical limitations of the approach are not well understood � Compilation (with propositional specifications) is equivalent to an NP-complete problem GAPPS RULER � The more expressive the agent specification language, the harder it is to compile it “The key lies in understanding how a process can � (There are some deep theoretical results naturally mirror in its states subtle conditions in its which say that after a certain expressiveness, environment and how these mirroring states ripple the compilation simply can’t be done.) out to overt actions that eventually achieve goals.” From Rosenschein and Kaelbling, “A Situated View of Representation and Control”, 1994 5-21 5-22 Advantages of Reactive Agents Limitations of Reactive Agents � Agents without environment models must have sufficient information available from local environment � Simplicity � If decisions are based on local environment, how does � Economy it take into account non-local information (i.e., it has a “short-term” view) � Computational tractability � Difficult to make reactive agents that learn � Robustness against failure � Since behavior emerges from component interactions � Elegance plus environment, it is hard to see how to engineer specific agents (no principled methodology exists) � It is hard to engineer agents with large numbers of behaviors (dynamics of interactions become too complex to understand) 5-23 5-24 4