only single action supported

[mattinjersey]

Getting this error...is this my mistake or do some agents only support a single action?

ValueError: Only a single action is supported by this network
In call to configurable 'CriticNetwork' (<function _NetworkMeta.new..capture_init at 0x00000203ED9DEEA0>)
In call to configurable 'train_eval' (<function train_eval at 0x00000203EE878BF8>)

[sguada]

You need to define a network that can support multiple Actions, you can use CriticNetwork as an example.

[mattinjersey]

Well for example with td3 agent, I define Critic Network as follows, and the action spec defined with 2 actions, and I get such error. The only agent that worked was PPOAgent.
Are we sure that all the agents can handle multiple actions?

_critic_net_input_specs = (tf_env.time_step_spec().observation,
tf_env.action_spec())

critic_net = critic_network.CriticNetwork(
    critic_net_input_specs,
    observation_fc_layer_params=critic_obs_fc_layers,
    action_fc_layer_params=critic_action_fc_layers,
    joint_fc_layer_params=critic_joint_fc_layers,
)_

[mattinjersey]

Right, I see in ddpg that critic_network can only handle 1 observation.
So it seems like ddpg agent is limited to 1 observation.

[mattinjersey]

SAC only limited to 1 observation, it appears.

[sguada]

So would need to write your own ActorNetwork and your own CriticNetwork classes that can handle multiple observations and multiple actions.

[mattinjersey]

ok I may try it. I'm not sure why the PPO Agent is written so that it can accept arbitrary number of actions/observations but the CriticNetwork is written differently.

[basvanopheusden]

I am also struggling with the same issue, @sguada, if you could give me some pointers as to how to create custom ActorDistributionNetworks, I would really appreciate it!