Online learning of multi-robot control
In this work we are interested in learning behaviors (navigating in an unknown environment, foraging food etc.) for a swarm of robots using Embodied Evolution.
- An animation showing a swarm of agents learning collaborative foraging.
PsyPhiNe is an interdisciplinary and exploratory project between philosophers, psychologists and computer scientists. The goal of the project is related to cognition and behavior. Cognition is a set of processes that are difficult to unite in a general definition. The project aims to explore the idea of assignments of intelligence or intentionality, assuming that our intersubjectivity and our natural tendency to anthropomorphize play a central role: we project onto others parts of our own cognition. To test these hypotheses, our aim is to design a "non-verbal" Turing Test, which satisfies the definitions of our various fields (psychology, philosophy, neuroscience and Computer Science), using a robotic prototype. some of the question that we aim to answer are: is it possible to give the illusion of cognition and or intelligence through such a technical device? How elaborate must be the control algorithms (behaviors) of such a device be to fool test subjects? How many degrees of freedom must it have?
This work is funded under the folowing grants:
2016 - 2017, A PostDoc Grant from the Lorraine Region.
2016 - 2018, Psyphine: Cogitamus ergo sumus a MSH & Université de Lorraine project.
2015 - 2016, PsyPhINe: Cogito Ergo Es a CNRS & Université de Lorraine PEPS Mirabelle project.
2015, an exploratory project grant from LORIA.
- A project report.
Evolutionary learning of Tetris controllers
The game of Tetris is probably among the few games that are popular worldwide and designing artificial players is a very challenging question. Our aim in this paper is to show that evolutionary computation can be applied to design good strategies for the game of Tetris that can compete with the best reported players. The player strategy uses a feature-based game evaluation function to evaluate the best moves for a given state and which is optimized using the CMA evolution strategy.
We have shown that by analyzing the shape of the search space, we were able to adapt the algorithm to converge better in this particular case. We also discuss how to reduce the length of the games which serves greatly when running lengthy learning algorithms. The racing procedures presented allow to learn very good player at a fraction of the learning cost.
We think that designing good artificial players should not rely only in optimizing the weights of a set of features, but also in choosing good feature functions. In fact our different experiments showed us that for the feature set presented here, we have probably obtained the best player and further improvements will not be significant if we take into account confidence intervals. Improvements could be achieved using better feature functions, ones that synthesize the game better. An interesting research question would be to devise algorithms to learn new feature functions rather than only learn the weights.