Current research

My research is focused in the topic of large-scale zero-sum games, for example: computer games of many genres (fighting, football, real-time strategy, etc.)

Those games have huge space-action spaces, where traditional planning and learning approaches are insufficient. Currently, I investigate novel approaches to such scenarios.

Initially, a task allocation approach was proposed to deal with the complexity of real-time strategy game StarCraft. For task allocation, we adopted Swarm-GAP, which is based on the behavior of social insects [1]. We used a genetic algorithm to adjust the task allocation parameters. The approach obtained promising results, and resulted in the best paper of SBGames 2014 Computing Track [2].

The task allocation approach is promising, but requires prior knowledge to encode a game's objective into tasks. Another approach to deal with large-scale games is to add a layer of reasoning above such games, so that we reason in a simplified representation.

For example, in real-time strategy games, several strategies are known. We can benefit by reasoning in the space of strategies. For instance, we can instantiate the process of strategy selection as a normal-form (matrix) game. Let's call it the metagame. The metagame's payoff matrix can be filled with the relative performance among strategies. Thus, if we calculate the metagame's Nash Equilibrium we obtain a strategy selection method with guaranteed expected performance.

Such approach resulted in a paper [3] and further investigation is being conducted.

References

[1] FERREIRA JR., P. R. Coordenacao de sistemas multiagente atuando em cenarios complexos: uma abordagem baseada na divisao de trabalho dos insetos sociais. 2008. PhD thesis - Universidade Federal do Rio Grande do Sul.

[2] Anderson R. Tavares, Hector Azpurua, and Luiz Chaimowicz. Evolving Swarm Intelligence for Task Allocation in a Real Time Strategy Game. In Computer Games and Digital Entertainment (SBGAMES), 2014 Brazilian Symposium on, pages 99-108, November 2014. [ bib | DOI | slides | http ]

[3] Anderson Tavares, Hector Azpúrua, Amanda Santos, and Luiz Chaimowicz. Rock, Paper, StarCraft: Strategy Selection in Real-Time Strategy Games. In Artificial Intelligence and Interactive Digital Entertainment (AIIDE), pages 93--99, October 2016. [ bib | slides | http ]

Past research

I also worked in distributed task allocation algorithms for complex scenarios. One benchmark for these scenarios is the Robocup Rescue (RCR) disaster simulator. In RCR, 3 teams of agents (ambulances, firefighters and police officers) must coordinate themselves in order to minimize the damage caused by an earthquake.

An idea for task allocation in complex scenarios

Currently, I'm not pursuing this research topic, but an interesting idea I was about to investigate was to try to capture human behavior complex scenarios. In order to do this, an interesting approach would involve building a computer game that embeds the charateristics of the RCR scenario and make it available for human players. The data recorded from the game would be used to (a) build an heuristic approach based on observation of human behavor and/or (b) use (or develop) machine learning techniques to extract patterns of human actions in the game and use them in the RCR virtual agents.