Monty.Frameworks.Models.MotorPolicies.PredefinedPolicy (ExTbpMonty.Core v0.1.0)

Copy Markdown View Source

A policy that replays a written-down sequence of actions, for ever.

Mirrors tbp/monty/frameworks/models/motor_policies.py::PredefinedPolicy. One action per line of a JSON file, read once at construction and then cycled: step n of the episode is action n of the file, wrapped round. It exits because unit tests and recorded-session experiments need a body that moves the same way twice, and no policy can promise that as simply as a list.

The file is read with ActionJSONDecoder, the same decoder an action crosses a wire with, so a file is Monty's own action serialisation and nothing bespoke.

Monty indexes with bounding % len(action_list), so a file with no actions fails on the first step with a division by zero. This port refuses at construction instead — same failure, one step earlier, and a message that says which file.

Summary

Types

t()

Functions

call(policy, ctx, observations, state, percept, goal)

The next action of the file, and the policy that has taken one more step.

new(options)

A replay for :agent_id from the actions written in :file.

read_action_file(file)

The actions in file, in the order they are written.

reset(policy)

Starts the next episode at the first action of the file.

state_dict(policy)

How far into the file the replay has got.

Types

t()

@type t() :: %Monty.Frameworks.Models.MotorPolicies.PredefinedPolicy{
  action_list: [Monty.Frameworks.Actions.Action.t()],
  agent_id: atom() | String.t(),
  episode_step: non_neg_integer()
}

Functions

call(policy, ctx, observations, state, percept, goal)

@spec call(
  t(),
  Monty.Context.t(),
  map() | nil,
  Monty.Frameworks.Models.MotorSystemState.t(),
  Monty.Cmp.Message.t() | nil,
  Monty.Cmp.Goal.t() | nil
) :: {Monty.Frameworks.Models.MotorPolicies.MotorPolicyResult.t(), t()}

The next action of the file, and the policy that has taken one more step.

The percept, the state, the observations and the goal are all ignored: the sequence is the whole of what this policy knows.

new(options)

@spec new(keyword()) :: t()

A replay for :agent_id from the actions written in :file.

:file is a path with one JSON action per line, blank lines allowed.

read_action_file(file)

@spec read_action_file(Path.t()) :: [Monty.Frameworks.Actions.Action.t()]

The actions in file, in the order they are written.

A blank line is not an action. A line that is not an action is an error naming the line, because a replay that silently skipped one would be a different experiment.

reset(policy)

@spec reset(t()) :: t()

Starts the next episode at the first action of the file.

state_dict(policy)

@spec state_dict(t()) :: map()

How far into the file the replay has got.