Probabilistic Selection in AgentSpeak(L)

TL;DR

The paper tackles autonomy under uncertainty by proposing a two-layer agent-programming framework that preserves symbolic BDI components (Jason/ASL) while incorporating probabilistic decision-making. It argues that probabilistic selection, via MEU and influence diagrams, can be integrated without rewriting the symbolic layer and demonstrates the approach conceptually using the GoldMiners scenario. The key contribution is a concrete method to bound probabilistic reasoning to the intention-selection function and generate a deliberation policy from a probabilistic graphical model that plugs into Jason. This hybrid architecture seeks to enhance robustness to sensor noise and dynamic environments while maintaining a clear separation of concerns between symbolic planning and probabilistic inference, with clear directions for real-time/off-line integration and learning in future work.

Abstract

Agent programming is mostly a symbolic discipline and, as such, draws little benefits from probabilistic areas as machine learning and graphical models. However, the greatest objective of agent research is the achievement of autonomy in dynamical and complex environments --- a goal that implies embracing uncertainty and therefore the entailed representations, algorithms and techniques. This paper proposes an innovative and conflict free two layer approach to agent programming that uses already established methods and tools from both symbolic and probabilistic artificial intelligence. Moreover, this framework is illustrated by means of a widely used agent programming example, GoldMiners.

Figures (3)

• Figure 1: The Jason deliberation process very resumed, with selection functions highlighted. The ASL only specifies the signature of the functions omitting any conditions, besides the type, on the output.
• Figure 2: If sensors report misreadings of the environment state the symbolic inference process inherent to BDI uses false perceptions as (true) facts of the environment and the deliberation process works on wrong assumptions. This plot relates sensor noise (the rate of sensor misreadings, in the horizontal axis) with agent performance measured by the the number of gathered golds (in the vertical axis). Two teams are plotted, the basic reference "dummy" that barely uses BDI features and the "smart" team, fully BDI, (designed by hubner2008developing) that won the 2006 "Multi-agent Programming Contest" dastani2007second featuring the GoldMiners scenario. Each data point summarizes the number of gathered golds by team in a given noise parameter and consists of the mean and standard variation of ten samples. The mean is traced by a thin black line and standard variation by a band centred in the mean value. Values between data-points are interpolations.
• Figure 3: An influence diagram for intuition selection of GoldMiners agents. The discovery, auctioning, transportation and deposit of gold pieces of the coordination protocol is represented in the graph. Variables (denoted by round rectangles) and Utilities (chamfered rectangles) in this diagram refer to total quantities of gold pieces. The probabilistic effect of the selected action is represented by the arrows that leave the action (rectangular) node.