Paul Davidsson

Towards anticipatory agents

This paper presents a novel approach to the problem of designing autonomous agents that is based on the idea of anticipatory systems. An anticipatory system has a model of itself and of the relevant part of its environment and will use this model to predict the future. The predictions are then utilised to determine the agents behaviour, i.e. it lets future states affect its present states. We argue that systems based on causal reasoning only, are too limited to serve as a proper base for designing autonomous agents. An anticipatory agent, on the other hand, will use reasoning from final cause to guide its current actions. We then discuss to what extent an anticipatory agent can be constructed from computable functions and we conclude that this problem is best expressed and analysed in linguistic terms. This discussion points out how such an agent should be designed and results in a proposal of an appropriate architecture. However, as the meta-linguistic problems involved are very hard to solve, a simpler architecture is also proposed. This is also a hybrid architecture that synthesizes reactive behaviour and deliberative reasoning, which, we believe, still has its merits compared to previous approaches to the design of autonomous agents. Finally, we discuss introspection and reflection, and show that the underlying concepts are easy to comprehend in the context of anticipatory systems.

A Linearly Quasi-Anticipatory Autonomous Agent Architecture: Some Preliminary Experiments

This report presents some initial results from simulations of a linear quasi-anticipatory autonomous agent architecture (Alqaaa), which correspond to a special case of a previously suggested general architecture of anticipatory agents. It integrates low-level reaction with high-level deliberation by embedding an ordinary reactive system based on situation-action rules, called the Reactor, in an anticipatory agent forming a layered hybrid architecture. By treating all agents in the domain (itself included) as reactive agents, this approach drastically reduces the amount of search needed while at the same time requiring only a small amount of heuristic domain knowledge. Instead it relies on a linear anticipation mechanism, carried out by the Anticipator, to achieve complex behaviours. The Anticipator uses a world model (in which the agent is represented only by the Reactor) to make a sequence of one-step predictions. After each step it checks whether the simulated Reactor has reached an undesired state. If this is the case it will modify the actual Reactor in order to avoid this state in the future. Results from both single- and multi-agent simulations indicate that the behaviour of Alqaaa agents is superior to that of the corresponding reactive agents. Some promising results on cooperation and coordination of teams of agents are also presented. In particular, the linear anticipation mechanism is successfully used for conflict detection.