Juhan Aru

Multiagent cooperation and competition with deep reinforcement learning

Evolution of cooperation and competition can appear when multiple adaptive agents share a biological, social, or technological niche. In the present work we study how cooperation and competition emerge between autonomous agents that learn by reinforcement while using only their raw visual input as the state representation. In particular, we extend the Deep Q-Learning framework to multiagent environments to investigate the interaction between two learning agents in the well-known video game Pong. By manipulating the classical rewarding scheme of Pong we show how competitive and collaborative behaviors emerge. We also describe the progression from competitive to collaborative behavior when the incentive to cooperate is increased. Finally we show how learning by playing against another adaptive agent, instead of against a hard-wired algorithm, results in more robust strategies. The present work shows that Deep Q-Networks can become a useful tool for studying decentralized learning of multiagent systems coping with high-dimensional environments.

Non-competitive liking for brands. No blocking in evaluative conditioning.

In the first experiment, we demonstrated evaluative conditioning using a novel across-modality procedure in which pictorial abstract brand logos acted as conditioned stimulus (CSs) and self-selected foods of different hedonic valence functioned as unconditioned stimuli (USs). We then investigated whether this form of learning of likes discriminates against redundant CSs using a blocking paradigm in the second experiment. The strength of evaluative conditioning accruing to the target CSs during compound training was unaffected by whether the other element of the compound was pretrained with a hedonic US. The observation that contingency learning about the target CS was blocked by the pretraining suggests that learning of likes and predictive learning are mediated by different processes.

On bounded-type thin local sets of the two-dimensional Gaussian free field

We study certain classes of local sets of the two-dimensional Gaussian free field (GFF) in a simply-connected domain, and their relation to the conformal loop ensemble CLE(4) and its variants. More specifically, we consider bounded-type thin local sets (BTLS), where thin means that the local set is small in size, and bounded-type means that the harmonic function describing the mean value of the field away from the local set is bounded by some deterministic constant. We show that a local set is a BTLS if and only if it is contained in some nested version of the CLE(4) carpet, and prove that all BTLS are necessarily connected to the boundary of the domain. We also construct all possible BTLS for which the corresponding harmonic function takes only two prescribed values and show that all these sets (and this includes the case of CLE(4)) are in fact measurable functions of the GFF.

Two-valued local sets of the 2D continuum Gaussian free field: connectivity, labels, and induced metrics

We study two-valued local sets,

Two Perspectives of the 2D Unit Area Quantum Sphere and Their Equivalence

\mathbb{A}_{-a,b}

Gaussian multiplicative chaos through the lens of the 2D Gaussian free field

, of the two-dimensional continuum Gaussian free field (GFF) with zero boundary condition in simply connected domains. Intuitively,

Approximating Liouville measure using local sets of the Gaussian free field

\mathbb{A}_{-a,b}

Balancing sums of random vectors

is the (random) set of points connected to the boundary by a path on which the values of the GFF remain in

First passage sets of the 2D continuum Gaussian free field

[-a,b]

Critical Liouville measure as a limit of subcritical measures

. For specific choices of the parameters