examples

Examples

This directory includes a number of working examples of Acme agents. These examples are not meant to be comprehensive and instead show a number of common use cases under which Acme agents can be applied.

Quickstart

We include a quick start tutorial to get you started. This takes you through the construction of the D4PG agent as an example, but the general structure should apply to any of the agents and environment combinations below.

Control Suite examples

These include a number of agents running on the dm_control suite of continuous control tasks:

• run_d4pg: a deterministic policy gradient (D4PG) agent which includes a determinstic policy and a distributional critic.
• run_dmpo: a maximum-a-posterior policy optimization agent which combines both a distributional critic and a stochastic policy.

Note that these examples require a MuJoCo license in order to run. See our tutorial for more details or see the dm_control repository for further information.

Behaviour Suite examples

We also include simple baselines across the bsuite set of tasks:

• run_dqn: an off-policy DQN examples; and
• run_impala: an (on-policy) Impala agent;

Finally we also include an example using demonstrations:

• run_dqfd: the DQfD agent running on hard-exploration tasks within bsuite (e.g. deep sea) using demonstration data; and
• run_bc: a behaviour cloining agent.

as well as a model-based example:

• run_mcts: the MCTS agent running on the task suite using either a simulator of the environment or a learned model.