Interacting Agent Based Systems Dean Petters 1. What is an agent? 2. Architectures for agents 3. Emailing agents 4. Computer games 5. Robotics 6. Sociological simulations 7. Psychological simulations
What is an agent? There is a spectrum of agency: from agents who take direct orders, to agents who have increasingly more independence and ultimately free agents Travel and estate agents Sports agents Politicians and chief executives Entrepreneurs Autonomy in agency
What is an agent? There are many views of what is or is not an agent. This talk takes the view that agents are not just: distributed/concurrent systems such as java threads Agents are: autonomous, for example synchronisation occurs at run time and agents are self-interested
What is an agent? - autonomy in agent simulations An agents autonomy within the bounds of a simulation might not include freedom to change their higher level goals. In a traffic simulation could agents choose to get out and walk?
Architectures for agents: Dennett s Kinds of Minds Darwinian Minds: No learning within the agent s/organisms lifetime adaptation by mutation/crossover and natural selection Alife agents and agents whose behaviour is directed by genetic algorithms may possess Darwinian Minds
Architectures for agents: Dennett s Kinds of Minds Skinnerian Minds: Associative learning/conditioning possible, Pavlov s dog and pigeon experiments Neural networks controllers Agents that learn by re-inforcement
Architectures for agents: Dennett s Kinds of Minds Popperian Minds: Good tests kill flawed theories; we remain alive to guess again. Deliberating/reasoning about what action to take before you take that action leaves you alive to deliberate again. Planning. Representing situations that you have never experienced.
Architectures for agents: Dennett s Kinds of Minds Gregorian Minds: Standing on the shoulders of giants. Using knowledge gained by others, (by reading, being lectured to etc), and therefore not having to rely upon your own experience, perceptions or reasoning Imitation, Sense of Self, Theory of mind
Architectures for agents: putting it all together How might we represent all the possibilities of any kind of mind? The cogaff-schema is a framework that allows us to systematically compare different architectures. An important issue is that between each category there are very many intermediate types.
Architectures for agents: deliberative, reactive and hybrid architectures Deliberative agents: - Are the dominant approach in (D)AI - May not react fast enough in dynamic environments - They have a well developed technology and methodology Reactive agents: - Are well represented in robotics and computer games - Often need to be custom built for each new problem Hybrid agents: - Can give us the best of both
Emailing agents An ideal environment for a reasoning agent. Stan Franklin - (Miami) -An agent that processes posting-transfer requests for the US army. Each soldier emails their request and the agent considers all requests attempting to satisfy as many individual and personnel requirements as possible. Because it is based on a theory of consciousness it is a real application and a simulation.
Computer games Not much AI in computer games. Many games require too much processing power for the graphics. Games developers prefer to add emotion by changing the way characters look rather than giving much more subtlety to the way characters behave.
Robotics Rodney Brooks and the COG projects. Reactive robots that use the subsumption architecture. Birmingham s COSY project. Not a solely reactive robot. Playmate scenarios that include the robot possessing the ability to talk about its actions.
Sociological simulations The EOS project: modelling paeleolithic social change - implemented in Prolog - inaccessible - appears non-deterministic - discrete (limited number of states) - episodic - static (not moment to moment) - multi-agent
Psychological simulations Robotics such as COG and the COSY project Drescher and simulating Piaget s theory of infant cognitive development. Simulating Bowlby s theory of Attachment theory (of social and emotional development) What advantage does using multi-agent simulations give us over other types of computational or mathematical modelling? - dynamic interaction - emergent properties - social analogy of co-evolution - outside in rather than inside out
Summary Multiple definitions. Frameworks for representing architectures. Applications. Simulations as scientific theories.