An Object Model for a Behavioural Planning in a Dynamic Multi-Agent System

Alex Whittaker and Tiziano Riolfo

A object oriented design template is described for the behavioural control
of a multi-agent system.

The template presented is being developed at Psygnosis Camden studio for
the game Buddies (working title). Broadly speaking, the game presents the
player with a team of up to six agents one of which is under their control,
the remainder being autonomous but responsive to commands.  This team is
pitted against up to seven other teams, which may be controlled by the
computer or by other players across a network.

The game world exists as a three dimensional terrain about which the agents
can move freely. The agents are anthropomorphic and their actions are
limited to movement in two dimensions, jumping, carrying crates, firing
weapons and driving vehicles. As well as static obstacles within the
terrain there are three other features of note:

Crates:        These enter the world in designated zones at a constant
rate, they represent the resource with which
          agents are able to build toys.
Stacking Pads: By delivering crates to the stacking pads, agents are able
to build larger crates which when hatched
          reveal toys whose value is broadly proportional to the number of
crates put into the pad.
Toys:          Revealed from the crates, toys can be weapons, vehicles and
new team members. Agents need to
          stack more crates in order to get more powerful toys, which will
increase their ability to win the game.

There are also neutral agents within the game world e.g. animals and
civilians, with which the player may interact.  These must also be
controlled with some degree of intelligence, bringing the maximum number of
agents in the game world to approximately sixty.

The vehicles in the game world mean that the agents must be able to use
different control systems depending on the vehicle type. Certain toys can
give the agents special abilities that should change the behaviour,
furthermore because there are several different types of buddies agents
will appear with different abilities, strengths and weaknesses.

The player must be able to control the agents in their team, however they
will expect them to behave intelligently without orders.  Because the
control interface must be kept quite simple, agents must be able to
interpret player instructions according to their condition and that of
their team.

Platform

A major constraint on the game design is the target hardware platform - a
Sony Playstation.  Whilst this represents a powerful tool for the
manipulation of graphical images, it is not an ideal platform on which to
tackle the large search spaces of classical AI.  The machine has 2Mb of
main memory, of which we might expect to be allocated 500Kb for data,
manipulated by a 33Mhz processor. If the system were to be compared to a
PC, the equivalent computing power would be a 486-generation processor with
a high-end graphics card and no floating-point operations.

Architecture

We have implemented a behavioural model using an augmented transition
network (ATN) at the highest level, a partial planner to execute some
operations such as route planing, and a model for representation of a
database of completed and partial plans.  The model can be extended to
allow a completely implemented partial planner for all actions, however we
have so far avoided this because of the restrictions of search.

The ATN is driven by percepts from the agents embodied in the game world.
Percepts are either member variables (The registers of the ATN) or
calculations taken from the game world from the agent's perspective e.g.
Agents health, distance from agent to nearest crate etc.

The ATN, the partial plan database and the percept database are described
in data files and generated in an editor that allows the agent behaviour to
be described by a designer rather than the programmer.

Performance

The system development is ongoing and the scheduled for completion to beta
level on the Playstation platform by summer 1999. We have implemented the
ATN on a PC platform and can demonstrate favourable performances in terms
of agent intelligence and reaction speed.

The general nature of the solution makes it applicable to a wide range of
problems where there are a large number of agents operating under a
constrained platform performance.

Presentation

We propose to present the object model, the behavioural editor and live
examples of the game, running on a PC platform and demonstrating the AI in
action.