CAQL: Continuous Action Q-Learning

Authors: Moonkyung Ryu, Yinlam Chow, Ross Anderson, Christian Tjandraatmadja, Craig Boutilier

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare CAQL with state-of-the-art RL algorithms on benchmark continuous-control problems that have different degrees of action constraints and show that CAQL outperforms policy-based methods in heavily constrained environments, often dramatically. ... We evaluate CAQL on one classical control benchmark (Pendulum) and five Mu Jo Co benchmarks (Hopper, Walker2D, Half Cheetah, Ant, Humanoid).
Researcher Affiliation Industry Moonkyung Ryu*, Yinlam Chow*, Ross Anderson, Christian Tjandraatmadja, Craig Boutilier Google Research {mkryu,yinlamchow,rander,ctjandra,cboutilier}@google.com
Pseudocode Yes Algorithm 1 Continuous Action Q-learning (CAQL)
Open Source Code No The paper does not provide a direct link or explicit statement about the open-sourcing of the CAQL implementation code developed in this paper.
Open Datasets Yes We evaluate CAQL on one classical control benchmark (Pendulum) and five Mu Jo Co benchmarks (Hopper, Walker2D, Half Cheetah, Ant, Humanoid). ... Table 6: Benchmark Environments.
Dataset Splits No The paper discusses training on data collected from simulation environments and stored in a replay buffer, sampling mini-batches for training. It does not specify fixed train/validation/test dataset splits in the conventional sense for static datasets.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU/CPU models or cloud instance types.
Software Dependencies Yes We use SCIP 6.0.0 (Gleixner et al., 2018) for the MIP solver.
Experiment Setup Yes Details on network architectures and hyperparameters are described in Appendix D. ... We use a two hidden layer neural network with Re LU activation (32 units in the first layer and 16 units in the second layer) for both the Q-function and the action function. ... A time limit of 60 seconds and a optimality gap limit of 10 4 are used for all experiments. For GA and CEM, a maximum iterations of 20 and a convergence threshold of 10 6 are used for all experiments if not stated otherwise. Table 7: Hyper parameters settings for CAQL and NAF. Table 8: Hyper parameters settings for DDPG, TD3, SAC.