Vinay Papudesi and Manfred Huber
INTRODUCTION

Staged skill learning involves:
 To
Begin:
 “Skills”
world.
 The
are innate reflexes and raw representation of the
Process:
 Abstract
away details of learnt skills
 Use these abstractions as part of a higher-level
representation:
Behavioural results
 Affordances

 Rinse
and repeat
THE DEVELOPMENTAL LEARNER

State representation encodes only those
aspects of the environmental state owing
behavioural and reward implications in the
context of its current capabilities.
A
compact representation
 Becomes more and more abstract over time

But how to model this?...
STATE-SPACES

Three yummy flavours:





External (World) State Space
Internal State Space
Action State Spaces
(…maps to…)
(…composed of…)
Internal and External spaces are good friends:
Where: Internal state
= Si
External state
= Se
Mapping function = I

Si ← I(Se)

Objective: Don’t hard-code mapping function, automate it!
Internal State Space is a vector of Action Spaces, one for each action
the agent provides…
ACTION SPACE
An action space is defined as a vector of paired
(indicator, predicator) conditions.
 Conditions are task-agnostic

Can be reused for learning different tasks
 Improvement over previous work


When an action is performed:
Signals a transition between internal states, S1 → S2.
 Observes an outcome from the world, oʹ.
 Two conditions are constructed:

 Indicator:
 Predicator:
Cind(S2) = oʹ
Cpre(S1) = oʹ
OUTCOMES, GENETIC ALGORITHMS, NON-DETERMINISM, OH MY!
World state space is potentially vast
 Must measure outcome somehow



Genetic Algorithms (GAs) are used to train hierarchical,
rule-based, classifiers
What if an outcome cannot be accurately
measured?
Classifiers simply flag world state as non-deterministic.
 Outcome is thus a triple type:
(success%, failure%, undetermined)

‘FIND’ ACTION
“Rotate 360° or until an object is visible”
TASKS
With the abstract state space constructed, the
agent can now learn optimal policies for
completing tasks.
 Treat the problem as a Markov Decision Process
(MDP).



From some internal state the agent must select an
appropriate action to progress toward completing the
task optimally.
Reinforcement learning is used to compute such
policies:
Select the policy which maximises the expected future
return.
 Future reward is estimated from prior experience.

THE TASK MODEL

Must acquire a Task Model
 Agent
interacts with environment, recording
experiences as it does so.
 The internal source and destination states get
updated with new conditions.
 The reward function is re-computed as the average
reinforcement value over all the recorded
experiences pertaining to the chosen action.
 Will
eventually converge on the true model
TASK-SPECIFIC CONDITIONS

Not all tasks can be optimally represented with
this approach.
Actions are individually encapsulated, knowledge
contained within them is not shared among them.
 E.g. ‘GOTO’ and ‘PICK’


Solution is to build ‘bipartition’ states
Allow the GOTO task a condition on whether the item
can be PICKed.
 … but only if the reward for doing so is significant and
the condition is statistically stable (low variance) and
deterministic.

RESULTS - FORAGING

Left:
A
hard-coded,
expert-designed
state space and
policy.

Right:
 Dynamically
acquired
equivalent.
RESULTS – STATE SPACE SIZE
As the agent
interacts with the
environment the
proposed
algorithm
maintains a nearconstant state
space complexity.
 The
representation is
continually
abstracted.

RESULTS – POLICY PERFORMANCE


The presented
technique is
comparable to
manually-designed
behaviour.
Domain specific
models are slow to
converge.

Their state spaces
are more complex
= harder to learn.
CONCLUSIONARY SENTIMENTS
The paper describes an approach that constructs
an abstract internal state space that is grounded
in the set of actions that the agent provides.
Reinforcement learning aids in selecting actions to
complete tasks.
 By applying an inherently epigenetic design they
have devised a developmental learner that
produces results that are comparable to handrolled solutions.
 Task learning is performed in a bottom-up fashion
(actions to tasks), but the representation of new
tasks thereafter can be constructed from the topdown using previously acquired state abstractions.
