In reinforcement learning (RL), it is easier to solve a task if given a good representation.

These RL algorithms are designed with a flat Markovian view of the environment in the form of a "state, action, reward, and next state" interface [

The inferred belief context can be leveraged to augment the state, leading to a policy that can adapt to abrupt variations in context.

As a challenging multi-player card game, DouDizhu has recently drawn much attention for analyzing competition and collaboration in imperfect-information games. In this paper, we propose PerfectDou, a state-of-the-art DouDizhu AI system that dominates the game, in an actor-critic framework with a proposed technique named perfect information distillation.

Our second contribution is to introduce an action semantic inductive bias and implement it with a technique named Entity-Bound Action Layer.

The green lines in Figure 1 demonstrate that the stiffness decreases as the number of training levels increases in most of the Procgen games. This suggests that the delayed critic update effectively alleviates the memorization problem. Each agent is trained on 200 training levels for 25M environment steps. Each agent is trained for 8M environment steps. The mean is computed over 10 different runs.

From the earliest years of our lives, humans use language to express our beliefs and desires. Being able to talk to artificial agents about our preferences would thus fulfill a central goal of value alignment. Y et today, we lack computational models explaining such language use.