There were five works in the session, one of them “Designing and Evolving an Unreal Tournament 2004 Expert Bot“.
It describes the designing and improvement, through off-line (not during the game) evolution, of an autonomous agent (or bot) for playing the game Unreal Tournament 2004. This was created by means of a finite state machine which models the expert behaviour of a human player in 1 vs 1 deathmatch mode, following the rules of the international competition.
Then, the bot was improved by means of a Genetic Algorithm, yielding an agent that is, in turn a very hard opponent for the medium-level human player and which can (easily) beat the default bots in the game, even in the maximum difficulty level.
The presentation can be seen at:
Moreover, you can watch one example of the evolution in the following video:
Recently, inside the last LION 7 (2013) conference (Special Session on Games and Computational Intelligence) there was presented the paper entitled “FSM-Based Agents for Playing Super Mario Game”.
It describes the implementation and test of an autonomous agent which can play Super Mario game better than an expert user can do (in some trained levels).
It is build starting from a Finite State Machine and applying an Evolutionary Algorithm.
The presentation is:
You can watch one example of the obtained agent playing a game here:
“Adaptative bots for real-time strategy game via map characterization” (A.Fernández-Ares, P.García-Sánchez, A.M. Mora, J.J Merelo) is the title of the paper we have presented in CIG2012. In this work we use Genetics Algorithms for improve an adaptative bot for play (and win!) to planet wars. We made it through the characterization of the maps, studing those features (calculated quickly) that influence in bot behavior:
This paper presents a proposal for a fast on-line map analysis for the RTS game Planet Wars in order to define specialized strategies for an autonomous bot. This analysis is used to tackle two constraints of the game, as featured in the Google AI Challenge 2010: the players cannot store any information from turn to turn, and there is a limited action time of just one second.They imply that the bot must analyze the game map quickly, to adapt its strategy during the game. Based in our previous work, in this paper we have evolved bots for different types of maps.
Then, all bots are combined in one, to choose the evolved strategy depending on the geographical configuration of the game in each turn. Several experiments have been conducted to test the new approach, which outperforms our previous version, based on an off-line general training.
This paper is a part of my Final Degree Project and it’s the result of our participation in the Google AI Contest of 2010. It’s also my first presentation in an conference, and the first time in English. In this paper we talk about the design of a bot that can play (and win) to the game Planet Wars. In this post we can read the rules of the contest and the game.
In this paper, we study the impact of the noisy fitness in the desing of the bot, because the choose of a bad fitness can make useless the genetic algorithm.
In this work, we present the results obtained from comparing several migration policies that tries to optimize in a noisy fitness environment: the on-line, on-board and hybrid evolutionary robotics problem. Three different migration policies have been studied (the most different migrant, random migrant and best migrant) and two replacement mechanisms: the migrant replaces the worst, or the migrant replaces the worst after being evaluated only if is better. Experiments with 4, 16 and 36 robots were conduced, with two different topologies (ring and panmictic) and also a comparison with other evolutionary robotics algorithms were performed. Results show that the replacement mechanism has more influence than the migration policy or topology, and it also affects the tuning of the algorithm parameters. We asked ourselves the next questions:
Using the hybrid approach (island model), which is the best combination of migration policy, admission policy, and island topology?
Is this combination better than the encapsulated and distributed alternatives?
Does the number of robots affect the result and if so, how?
Conclusions, graphs and stuff and in the paper, but summarizing, multikulti technique (receive the most different individual of my population from other islands) and accept it in my population after its evaluation perform better than other alternatives, even with less migration rate.
We investigate on-line on-board evolution of robot controllers based on the so-called hybrid approach (island-based). Inherently to this approach each robot hosts a population (island) of evolving controllers and exchanges controllers with other robots at certain times. We compare different exchange (migration) policies in order to optimize this evolutionary system and compare the best hybrid setup with the encapsulated and distributed alternatives. We conclude that adding a difference-based migrant selection scheme increases the performance.
Tomorrow we will be presenting the work “Validating a Peer-to-Peer Evolutionary Algorithm” in Evo* 2012 held in Malaga, Spain. You can find below the abstract and presentation of the work.
This paper proposes a simple experiment for validating a Peer-to-Peer Evolutionary Algorithm in a real computing infrastructure in order to verify that results meet those obtained by simulations. The validation method consists of conducting a well-characterized experiment in a large computer cluster of up to a number of processors equal to the population estimated by the simulator. We argue that the validation stage is usually missing in the design of large-scale distributed meta-heuristics given the difficulty of harnessing a large number of computing resources. That way, most of the approaches in the literature focus on studying the model viability throughout a simulation-driven experimentation. However, simulations assume idealistic conditions that can influence the algorithmic performance and bias results when conducted in a real platform. Therefore, we aim at validating simulations by running a real version of the algorithm. Results show that the algorithmic performance is rather accurate to the predicted one whilst times-to-solutions can be drastically decreased when compared to the estimation of a sequential run.
Aquí está la presentación del trabajo que da título al post. Es una versión actualizada del trabajo que se presentó en el IWANN 2011, así que os refiero primero a esa versión por si no estáis al tanto.
For information about an early version of this work (in English) please check here.
Our studies on the sandpile mutation operator were extended, presented at MAEB2012 and published in the proceedings (in spanish). The abstract:
La mutación montón de arena es un operador para Algoritmos Genéticos basado en un modelo de Criticalidad Auto-Organizado con el mismo nombre. El operador ha sido desarrollado con el objetivo de resolver problemas con función objetivo variable. Este artículo propone un estudio del operador y la optimización de su desempeño, experimentando diferentes estrategias que conectan el modelo auto-organizado y el Algoritmo Genético. Las pruebas sobre el algoritmo se desarrollan con un gran conjunto de problemas dinámicos, diseñados con un generador de problemas a partir de funciones-base no dinámicas. Las mejores configuraciones del algoritmo son comparadas con dos AGs recientemente propuestos para optimización dinámica. Demostramos que un AG con el operador de mutación montón de arena es eficiente para el conjunto de pruebas que es propuesto.