Emergence of Implicit World Models from Mortal Agents

We discuss the possibility of world models and active exploration as emergent properties of open-ended behavior optimization in autonomous agents. In discussing the source of the open-endedness of living things, we start from the perspective of biological systems as understood by the mechanistic approach of theoretical biology and artificial life. From this perspective, we discuss the potential of homeostasis in particular as an open-ended objective for autonomous agents and as a general, integrative extrinsic motivation. We then discuss the possibility of implicitly acquiring a world model and active exploration through the internal dynamics of a network, and a hypothetical architecture for this, by combining meta-reinforcement learning, which assumes domain adaptation as a system that achieves robust homeostasis.

On Reward Function for Survival

Obtaining a survival strategy (policy) is one of the fundamental problems of biological agents. In this paper, we generalize the formulation of previous research related to the survival of an agent and we formulate the survival problem as a maximization of the multi-step survival probability in future time steps. We introduce a method for converting the maximization of multi-step survival probability into a classical reinforcement learning problem. Using this conversion, the reward function (negative temporal cost function) is expressed as the log of the temporal survival probability. And we show that the objective function of the reinforcement learning in this sense is proportional to the variational lower bound of the original problem. Finally, We empirically demonstrate that the agent learns survival behavior by using the reward function introduced in this paper.

Q-Networks for Binary Vector Actions

In this paper reinforcement learning with binary vector actions was investigated. We suggest an effective architecture of the neural networks for approximating an action-value function with binary vector actions. The proposed architecture approximates the action-value function by a linear function with respect to the action vector, but is still non-linear with respect to the state input. We show that this approximation method enables the efficient calculation of greedy action selection and softmax action selection. Using this architecture, we suggest an online algorithm based on Q-learning. The empirical results in the grid world and the blocker task suggest that our approximation architecture would be effective for the RL problems with large discrete action sets.