The dynamics of teams can be investigated by agent based modelling where team behaviour emerges from the actions and interactions of individual team members as autonomous agents. Let xi be an individual agent with state st(xi) at time t. Let the set y = <x1, …, xn> be the set of agents in the team. Let st(y) be the state of the team at time t. Let the changes in state st1(xi) → st2(xi) and st1(y) → st2(y) be called events. et2t1(xi) = st2(xi) – st1(xi) is Level-1 event and et2t1(y) = st2(yj) – st1(yj) is Level-2 event. Together et2t1(xi), i = 1, …, n and et2t1(yj) form a two-level event with coupled sub-events at Levels 1 and 2. When teams <y1, … , ym> interact they form higher level structures, z, with higher level events at Level 3. Then we have multilevel events with coupled sub-events at Levels 1, 2 and 3. The two-level dynamics of heterogeneous social systems with bottom-up and top-down coupling can make them so complex that agent-based simulation is the only known way to explore their dynamics. The extension of this to multilevel systems with many intermediate levels and multilevel patterns through time remains a research challenge. Robot soccer teams have the same structural properties as human teams but are much easier to study, and provide a laboratory for initial investigations. Figure 1(a) illustrates a common emergent structure in soccer games that can be represented as <r1, r2, b1, b2, ball, Gb; Rdd> with the elements combined by the defender’s dilemma relation Rdd. Figure 1(b) shows a subsequent structure <r1, r2, b1, b2, ball, Gb; Rgd> where the same players are now combined into a different structure under the goalkeeper’s dilemma relation, Rgd. The transition between these two states is an event, as is the transition to <r1, r2, b1, b2, ball, Gb; Rgs> where Rgs is the ‘goal scored’ relation with the ball in the goal Gb. Generally the instantaneous state of the system is more complex than this with many relations between subsets of the agents. Figure 2 shows a sequence of ball positions as a green path with the particular player positions and closeness relations at t = 4217. This can be considered to be a multilevel trajectory of structures through time with final event being a goal scored. The number of dynamic combinatorial multilevel structures is potentially astronomic, but within them are those that are precursors to improved position such a dominance of the pitch and goals. In our presentation we will show how the extensive data from many RoboCup soccer competition can be analysed using the methods sketched here to inform tactics and strategy in robot soccer and, more generally, to understand better team dynamics in multilevel social systems.
