Applications
of
Reinforcement Learning in the Power Systems Industry
(Invited Talk at the First Annual Power and Energy Forum, Nov
6, 2008)
Critterbot
Project Overview (Aug 6, 2008)
Mind and Time: A View of Constructivist
Reinforcement Learning
This an invited talk I gave at
the European Workshop on Reinforcement Learning in summer
2008. The basic idea is that in order to learn fast it
is necessary to learn slow, that the key to fast reinforcement
learning is to prepare for it by a slow continual process of
constructing a model of the world's state and dynamics.
Although I don't know exactly how to do this, I have many
ideas and suggestions, and an outline of how to proceed.
I try to communicate these in this talk.
How simple can mind
be? (International Workshop on Natural and Artificial
Cognition, University of Oxford 6/26/07)
On the Role of
Tracking in Stationary Environments (ICML'07 6/21/07) Associated
paper.
ABSTRACT: It is often thought
that learning algorithms that track the best solution, as
opposed to converging to it, are important only on
nonstationary problems. We present three results suggesting
that this is not so. First we illustrate in a simple concrete
example, the Black and White problem, that tracking can
perform better than any converging algorithm on a stationary
problem. Second, we show the same point on a larger, more
realistic problem, an application of temporal-difference
learning to computer Go. Our third result suggests that
tracking in stationary problems could be important for
meta-learning research (e.g., learning to learn, feature
selection, transfer). We apply a meta-learning algorithm for
step-size adaptation, IDBD,e to the Black and White problem,
showing that meta-learning has a dramatic long-term effect on
performance whereas, on an analogous converging problem,
meta-learning has only a small second-order effect. This small
result suggests a way of eventually overcoming a major
obstacle to meta-learning research: the lack of an independent
methodology for task selection.
Stimulus
Representation in Temporal-Difference Models of the Dopamine
System (Cal Tech 6/4/07)
The neurotransmitter dopamine
plays an important role in the processing of reward-related
information in the brain. A prominent theory of this function
is that the phasic firing of dopamine neurons encodes a reward
prediction error as formalized by the temporal-difference (TD)
algorithm in reinforcement learning. Most of these TD models
of the dopamine system have assumed a "complete serial
compound" representation in which every moment within a trial
is represented distinctly with no similarity to neighboring
moments. In this paper we present a more realistic temporal
representation in which external stimuli spawn a series of
internal microstimuli which grow weaker and more diffuse over
time. We show that if these microstimuli are used as inputs to
the TD model, then its match to experimental data is improved
for hitherto problematic cases in which reward is omitted or
received early. We also note that the new model never
produces large negative errors, suggesting that a second
neurotransmitter for representing negative errors may not be
necessary. Generally, we conclude that choosing a stimulus
representation with a more realistic temporal profile can
significantly alter the predictions of the TD model of
dopamine function.
Experience-Oriented Artificial
Intelligence (Machine Learning Seminar at the University
of Toronto, 4/3/06)
If intelligence is a
computation, then the temporal stream of sensations is its
input, and the temporal stream of actions is its output. These
two intermingled time series make up experience. They
are the basis on which all intelligent decisions are made and
the basis on which those decisions are judged. A focus on
experience has implications for many aspects of AI; in this
talk we consider its implications for knowledge
representation. I propose that it is possible and desirable
for an AI agent's knowledge of the world to be expressed
entirely as predictions about its low-level experience. Even
abstract concepts, such as the concept of a chair, can be
expressed as predictions, e.g., about what will happen if we
try to sit. The predictive approach is appealing because it
connects knowledge directly to data, allowing knowledge to be
autonomously verified and tuned, perhaps even learned.
However, there is a tremendous gap between human-level
knowledge (e.g., about space, objects, people, or water) and
low-level experience. The purpose of this talk is to
present some recent work suggesting how this gap might someday
be bridged. I describe a series of small experiments in
which extensions of reinforcement learning methods are used to
learn predictive representations of abstract commonsense
knowledge in micro-worlds. These are first steps on a long
journey toward understanding how a mind might make sense of
the blooming, buzzing confusion of its sensori-motor
experience.
Predictive Representations of State and
Knowledge (ICML'05 workshop on Rich Representations for
Reinforcement Learning, 8/7/05)
What is knowledge? The
empiricist answer, dating back to the 19th century, is that
knowledge is the ability to predict. In a modern version of
this idea, reinforcement learning researchers have proposed
that artificial agents should represent their knowledge as
predictions of their low-level sensations and actions.
This predictive representations (PR) approach is appealing
because it connects knowledge directly to data, thereby
facilitating learning and clarifying semantics. Most PR
research has emphasized representing the world's _state.
In this talk I will survey the main results and mathematical
ideas of that work. A natural follow on, just beginning to be
explored, is to use PRs for all kinds of world knowledge, of
dynamics as well as of state, of abstractions as well as
specifics. I will survey this work as well and attempt
to make vivid the potential of PRs for artificial
intelligence.
Grounding knowledge in subjective
experience (provocative remarks at the 2nd Cognitive
Systems Conference, 5/20/05)
Experience-Oriented Artificial
Intelligence (McGill 11/30/05)
I propose that experience - the
explicit sequence of actions and sensations over an agent's
life - should play a central role in all aspects of artificial
intelligence. In particular:
1. Knowledge representation
should be in terms of experience. Recent work has shown that a
surprisingly wide range of world knowledge can be expressed as
predictions of experience, enabling it to be automatically
verified and tuned, and grounding its meaning in data rather
than in human understanding.
2. Planning/reasoning should be in terms of experience. It is
natural to think of planning as comparing alternative future
experiences. General methods, such as dynamic programming, can
be used to plan using knowledge expressed in the
aforementioned predictive form.
3. State representation should be in terms of experience.
Rather than talk about objects and their metric or even
topological relationships, we represent states by the
predictions that can be made from them. For example, the state
"John is in the coffee room" corresponds to the prediction
that going to the coffee room will produce the sight of John.
Much here has yet to be worked
out. Each of the "should"s above can also be read as a
"could", or even a "perhaps could". I am optimistic and
enthusiastic because of the potential for developing a compact
and powerful theory of AI in the long run, and for many easy
experimental tests in the short run.
Grounding Commonsense Knowledge in
Question Networks. (University of Michigan 9/28/04)
A long-standing challenge in
artificial intelligence has been to relate the kind of
commonsense knowledge that people have about the world (for
example, about space, objects, people, trees and water) to the
low-level stream of sensations and actions. In this
talk, we present new work that brings us a few steps closer to
realizing this goal. We introduce the idea of question
networks, a way of expressing arbitrary machine-readable
questions about future sensations and actions, and a
temporal-difference algorithm for learning answers to the
questions. In a series of small experiments, we
illustrate the learning efficency of these methods and their
ability to handle non-Markov problems. Finally, we
present their extension to temporally abstract knowledge in
terms of closed-loop macro-actions known as options.
Overall, we argue that these steps bring us qualitatively
closer to understanding the blooming, buzzing confusion of
sensori-motor experience.
Temporal Difference Networks.
Presented at NIPS-04. Larger Version.
We introduce a generalization of
temporal-difference (TD) learning to networks of interrelated
predictions. Rather than relating a single prediction to
itself at a later time, as in conventional TD methods, a TD
network relates each prediction in a set of predictions to
other predictions in the set at a later time. TD networks can
represent and apply TD learning to a much wider class of
predictions than has previously been possible. Using a
random-walk example, we show that these networks can be used
to learn to predict by a fixed interval, which is not possible
with conventional TD methods. Secondly, we show that when
actions are introduced, and the inter-prediction relationships
made contingent on them, the usual learning-efficiency
advantage of TD methods over Monte Carlo (supervised learning)
methods becomes particularly pronounced. Thirdly, we
demonstrate that TD networks can learn predictive state
representations that enable exact solution of a non-Markov
problem. A very broad range of inter-predictive temporal
relationships can be expressed in these networks. Overall we
argue that TD networks represent a substantial extension of
the abilities of TD methods and bring us closer to the goal of
representing world knowledge in entirely predictive, grounded
terms.
Knowledge Representation in TD
Networks (AAAI Symposium on MDPs and POMDPs: Advances and
Challenges (7/26/04) Large (1024 x
768) version
We introduce a generalization of
temporal-difference (TD) learning to networks of interrelated
predictions. Rather than relating a single prediction to
itself at a later time, as in conventional TD methods, a TD
network relates each prediction in a set of predictions to
other predictions in the set at a later time. TD networks can
represent and apply TD learning to a much wider class of
predictions than has previously been possible. Using a
random-walk example, we show that these networks can be used
to learn to predict by a fixed interval, which is not possible
with conventional TD methods. Secondly, we show that when
actions are introduced, and the inter-prediction relationships
made contingent on them, the usual learning-efficiency
advantage of TD methods over Monte Carlo (supervised learning)
methods becomes particularly pronounced. Thirdly, we
demonstrate that TD networks can learn predictive state
representations that enable exact solution of a non-Markov
problem. A very broad range of inter-predictive temporal
relationships can be expressed in these networks. Overall we
argue that TD networks represent a substantial extension of
the abilities of TD methods and bring us closer to the goal of
representing world knowledge in entirely predictive, grounded
terms.
Toward a Computational Theory of
Intelligence -- iCORE talk on Reinforcement Learning and
Artificial Intelligence (University of Calgary 2/25/04).
Video no longer available.
This talk was to a general university audience (videocast to
U. Alberta and U. Lethbridge). To showcase the ideas and power
of RL, i collected a bunch of videos from other peoples' work.
It's not often you can do this appropriately, but I think it
was ok this time, and certainly it was fun. The accompanying
videos:
All save the last are
quicktimable and will play directly in safari. The last seems
to require mplayer.
Adapting bias by gradient descent: An
incremental version of delta-bar-delta (University of
Alberta 2/2/04)
Appropriate bias is widely viewed as the key to efficient
learning and generalization. I present a new algorithm, the
Incremental Delta-Bar-Delta (IDBD) algorithm, for the learning
of appropriate biases based on previous learning experience.
The IDBD algorithm is developed for the case of a simple,
linear learning system---the LMS or delta rule with a separate
learning-rate parameter for each input. The IDBD algorithm
adjusts the learning-rate parameters, which are an important
form of bias for this system. Because bias in this approach is
adapted based on previous learning experience, the appropriate
testbeds are drifting or non-stationary learning tasks. For
particular tasks of this type, I show that the IDBD algorithm
performs better than ordinary LMS and in fact finds the
optimal learning rates. The IDBD algorithm extends and
improves over prior work by Jacobs and by me in that it is
fully incremental and has only a single free parameter. This
paper also extends previous work by presenting a derivation of
the IDBD algorithm as gradient descent in the space of
learning-rate parameters. Finally, I offer a novel
interpretation of the IDBD algorithm as an incremental form of
hold-one-out cross validation.
From Markov Decision Processes to
Artificial Intelligence (University of Alberta 5/14/03)
The path to general, human-level intelligence may go through
Markov decision processes (MDPs), a discrete-time,
probabilistic formulation of sequential decision problems in
terms of states, actions, and rewards. Developed in the 1950s,
MDPs were extensively explored and applied in operations
research and engineering before coming to the attention of
artificial intelligence researchers about 15 years ago. Much
of the new interest has come from the field of reinforcement
learning, where novel twists on classical dynamic programming
methods have enabled the solution of more and vastly larger
problems, such as backgammon (Tesauro, 1995) and elevator
control (Crites and Barto, 1996). Despite remaining technical
issues, real progress seems to have been made toward general
learning and planning methods relevant to artificial
intelligence. We suggest that the MDP framework can be
extended further, to the threshold of human-level
intelligence, by abstracting and generalizing each of its
three components - actions, states, and rewards. We briefly
survey recent work on temporally abstract actions (Precup,
2000; Parr, 1998), predictive representations of state
(Littman et al., 2002), and non-reward subgoals (Sutton,
Precup & Singh, 1998) to make this suggestion.
The reinforcement learning
approach to understanding intelligence is now about 20 years
old, which should be time enough for a mature perspective on
what it is and what it has contributed. Reinforcement learning
methods, particularly temporal-difference learning, have been
widely used in control and robotics applications, in playing
games such as chess and backgammon, in operations research,
and as models of animal learning and neural reward systems.
Holding these diverse applications together, and posing as a
fundamental statement about cognition and decision-making, is
a computational theory (in the sense of Marr) of mind.
Reinforcement learning methods are centered around the
interaction and simultaneous evolution of two primary
functional objects, the policy, which says what to do in each
situation, and the value function, which says how desirable it
is to be in each situation. In this talk, I will survey
several examples of reinforcement learning in the attempt to
make this underlying theory vivid. Finally, I will mention
some of the theory's limitations and shortcomings, and ongoing
efforts to make it relevant to the extremely powerful and
flexible cognition that we see in humans.
Experience-Oriented
Artificial Intelligence (Nov 2002)
I propose that experience - the
explicit sequence of actions and sensations over an agent's
life - should play a central role in all aspects of artificial
intelligence. In particular:
1. Knowledge representation
should be in terms of experience. Recent work has shown that
a surprisingly wide range of world knowledge can be
expressed as predictions of experience, enabling it to be
automatically verified and tuned, and grounding its meaning
in data rather than in human understanding.
2. Planning/reasoning should be in terms of experience. It
is natural to think of planning as comparing alternative
future experiences. General methods, such as dynamic
programming, can be used to plan using knowledge expressed
in the aforementioned predictive form.
3. State representation should be in terms of experience.
Rather than talk about objects and their metric or even
topological relationships, we represent states by the
predictions that can be made from them. For example, the
state "John is in the coffee room" corresponds to the
prediction that going to the coffee room will produce the
sight of John.
Much here has yet to be worked out. Each of the "should"s
above can also be read as a "could", or even a "perhaps
could". I am optimistic and enthusiastic because of the
potential for developing a compact and powerful theory of AI
in the long run, and for many easy experimental tests in the
short run.
[some of this is joint work with Doina Precup, Michael
Littman, Satinder Singh & Peter Stone]
Artificial Intelligence Should Be
About Predictions (AT&T 12/7/01)
What keeps the knowledge in an
AI system correct? Usually people do, but that is a dead end;
eventually the AI must do it itself. Building AIs that can
maintain their own knowledge is probably the greatest single
challenge facing AI today.
It would be relatively easy to self-maintain knowledge if it
were expressed as predictions: you would predict something and
then see what actually happened. In this talk I propose that
much of our knowledge of the world can be expressed as
predictions that can be verified in this way. Certainly much
of our everyday decision-making is based on predictions about
alternative alternative courses of action. Even abstract
concepts, such as the concept of a chair, can be expressed as
predictions, e.g., about what would happen if we try to sit.
Emphasizing ideas rather than technical details, I will
describe some of the challenges to this predictive view and
partial solutions. The main challenge is to be able to express
in predictive form the wide variety of knowledge we have of
the world. This can be done in large part by allowing the
predictions to be conditional on action and to terminate
flexibly, as in the "options" framework. A second challenge is
to be fully grounded, to relate the meaning of predictions
directly to data. Finally, we consider the pragmatic
challenges: how to make progress with these ideas? Building a
self-maintaining AI based on predictive knowledge is not
difficult, but requires new ways of thinking, determination to
do it right, and a willingness to proceed slowly.
Mind is About Predictions
(Northeastern 7/31/01)
In this talk I will describe
recent research in artificial intelligence which has given
greater credance to the old idea that much of our knowledge of
the world is in the form of predictions. From the blooming,
buzzing confusion we extract what is predictable, and in so
doing discover useful concepts and ways of behaving.
Certainly, much of our everyday reasoning and decision making
is based on predictions about alternative courses of action.
Even abstract concepts, such as the concept of a chair, can be
expressed as predictions, e.g., about what will happen if we
try to sit. In this talk I will briefly cover three ideas: 1)
an expanded notion of prediction capable of expressing a broad
range of knowledge, 2) a kind of planning, or reasoning, as
the combination of predictions to yield new predictions, and
3) a way of representing the state of the world (as well as
its dynamics) as predictions. All this suggests that working
with predictions is what the mind is all about---that
predictions are the coin of the mental realm.
(Some of the newer bits of this are joint work with Michael
Littman, Doina Precup, and Satinder Singh; also many thanks to
David McAllester for constructive criticism.)
Off-policy temporal-difference
learning with function approximation (ICML 7/1/01)
We introduce the first
algorithm for off-policy temporal-difference learning that is
stable with linear function approximation. Off-policy learning
is of interest because it forms the basis for popular
reinforcement learning methods such as Q-learning, which has
been known to diverge with linear function approximation, and
because it is critical to the practical utility of
multi-scale, multi-goal, learning frameworks such as options,
HAMs, and MAXQ. Our new algorithm combines TD(lambda) over
state-action pairs with importance sampling ideas from our
previous work. We prove that, given training under any
epsilon-soft policy, the algorithm converges w.p.1 to a close
approximation (as in Tsitsiklis and Van Roy, 1997; Tadic,
2001) to the action-value function for an arbitrary target
policy. Variations of the algorithm designed to reduce
variance introduce additional bias but are also guaranteed
convergent. We also illustrate our method empirically on a
small policy evaluation problem, showing reduced variance
compared to the most obvious importance sampling algorithm for
this problem. Our current results are limited to episodic
tasks with episodes of bounded length.
Overcoming the Curse of Dimensionality
with Reinforcement Learning (MIT ORC 4/19/01)
Technological advances in the
last few decades have made computation and memory vastly
cheaper and thus available in massive quantities. The field of
reinforcement learning attempts to take advantage of this
trend when solving large-scale stochastic optimal control
problems. Dynamic programming can solve small instances of
such problems, but suffers from Bellman's "curse of
dimensionality," the tendency of the state space and thus
computational complexity to scale exponentially with the
number of state variables (and thus to quickly exceed even the
"massive" computational resources now available).
Reinforcement learning brings in two new techniques: 1)
parametric approximation of the value function, and 2)
sampling of state trajectories (rather than sweeps through the
state space). These enable finding approximate solutions,
improving in quality with the available computational
resources, on problems too large to even be attempted with
conventional dynamic programming. However, these techniques
also complicate theory, and there remain substantial gaps
between the reinforcement learning methods proven effective
and those that appear most effective in practice. In this
talk, I present results extending the convergence result of
Tsitsiklis and Van Roy for on-policy evaluation with linear
function approximation to the off-policy case, reviving the
possibility of convergence results for value-based off-policy
control methods such as Q-learning. I also present an
application to RoboCup soccer illustrating the linear approach
to function approximation. (This is joint work with Doina
Precup, Satinder Singh, Peter Stone, and Sanjoy Dasgupta.)
The Right Way to do
Reinforcement Learning with Function Approximation
(NIPS'00 12/2/00)
From Reflex to Reason
(Cornell 12/8/00)
How close are we to a
computational understanding of the mind? Perhaps closer than
is usually thought. In this talk I discuss a small set of
principles drawn from reinforcement learning and other parts
of artificial intelligence that cover a broad range of mental
phenomena, from reflexes through various kinds of learning,
planning, and reasoning. These principles include rewards,
value functions, state-space search, and, as I emphasize in
this talk, representing our knowledge of the world as
predictions of future observations. First, I show how
predictive representations provide a new theory of that
simplest of learning phenomena, Pavlovian conditioning or the
learning of replexes. Second, I briefly outline how
model-based reinforcement learning with mental simulation can
serve as a theory of reasoning. I argue that representing
knowledge as predictions, including the possibility of
action-contingent and temporally indefinite predictions,
solves critical problems in the semantics and grounding of
classical symbolic approaches to knowledge representation.
Toward Grounding Knowledge in
Prediction (CEC2000 7/18/00)
Any attempt to build
intelligent machines must come to grips with the question of
knowledge, of what kind of information about the world the
machine stores and manipulates. Traditionally there have been
two approaches, the horns of a dilemma. One uses verbal
statements like "John loves Mary" or "Socrates is a man" whose
meaning is clear only to people, not to machines; such
knowledge is ungrounded. The other uses mathematical
statements like differential equations or transition matrices
which, although clear and grounded, have never seemed adequate
for expressing the commonsense knowledge we all have about the
world and use everyday. In this talk we suggest that this
dilemma can be broken by grounding knowledge in an enlarged
notion of conditional prediction. In particular, if we allow
predictions conditional on outcomes (as in Precup, 2000; Parr,
1999) then much more can be expressed as predictions without
losing grounding and mathematical clarity. In addition, this
approach suggests a radical theory of reasoning---combining
knowledge to yield new knowledge---as simple composition of
predictions.
A Least Common Denominator for Temporal
Abstraction in Reinforcement Learning (NIPS workshop
12/5/98)
Improved
Switching Among Temporally Abstract Actions (NIPS 12/2/98)
In robotics and other control
applications it is commonplace to have a pre-existing set of
controllers for solving subtasks, perhaps hand-crafted or
previously learned or planned, and still face a difficult
problem of how to choose and switch among the controllers to
solve an overall task as well as possible. In this paper we
present a framework based on Markov decision processes and
semi-Markov decision processes for phrasing this problem, a
basic theorem regarding the improvement in performance that
can be obtained by switching flexibly between given
controllers, and example applications of the theorem. In
particular, we show how an agent can plan with these
high-level controllers and then use the results of such
planning to find an even better plan, by modifying the
existing controllers, with negligible additional cost and no
re-planning. In one of our examples, the complexity of the
problem is reduced from 24 billion state-action pairs to less
than a million state-controller pairs.
Reinforcement Learning: How Far Can
It Go? (Past, Present, and Future) (ICML/COLT/UAI 7/25/98,
Extended abstract)
Between MDPs and Semi-MDPs
(Stanford 3/5/98)
A key challenge for AI is how to
learn, plan, and represent knowledge at multiple levels of
temporal abstraction. In this talk I develop an approach based
on the mathematical framework of reinforcement learning and
Markov decision processes (MDPs). The usual framework is
extended to include closed-loop multi-step options---whole
courses of behavior that may be temporally extended,
stochastic, and contingent on events. Examples of options
include picking up an object, going to lunch, and traveling to
a distant city, as well as primitive actions such as muscle
twitches and joint torques. Options can be used
interchangeably with primitive actions in reinforcement
learning and planning methods, and can be analyzed in terms of
a generalized kind of MDP known as a semi-Markov decision
process (SMDP) (e.g., Puterman, 1994; Bradtke and Duff, 1995;
Parr, 1998; Precup and Sutton, 1997). In this talk I focus on
the interplay between the MDP and SMDP levels of analysis. I
show how a set of options can be improved by changing their
termination conditions to improve over SMDP planning methods
with no additional cost. I also present novel intra-option
temporal-difference methods that substantially improve
over SMDP methods. Finally, I discuss how options themselves
can be learned, introducing a new notion of subgoal and
subtask into reinforcement learning. Overall, I argue that
options and models of options provide hitherto missing aspects
of a powerful, clear, and expressive framework for
representing and organizing knowledge. (Joint work with Doina
Precup and Satinder Singh.)
Reinforcement Learning: A
Tutorial (GP-98 7/23/98)
Reinforcement learning is
learning about, from, and while interacting with a environment
in order to achieve a goal. In other words, it is a relatively
direct model of the learning that people and animals do in
their normal lives. In the last two decades, this age-old
problem has come to be much better understood by integrating
ideas from psychology, optimal control, artificial neural
networks, and artificial intelligence. New methods and
combinations of methods have enabled much better solutions to
large-scale applications than had been possible by all other
means. This tutorial will provide a top-down introduction to
the field, covering Markov decision processes and approximate
value functions as the formulation of the problem, and dynamic
programming, temporal-difference learning, and Monte Carlo
methods as the principal solution methods. The role of neural
networks and planning will also be covered. The emphasis will
be on understanding the capabilities and appropriate role of
each of class of methods within in an integrated system for
learning and decision making
Reinforcement Learning: Lessons
for Artificial Intelligence (IJCAI 8/28/97)
The field of reinforcement
learning has recently produced world-class applications and,
as we survey in this talk, scientific insights that may be
relevant to all of AI. In my view, the main things that we
have learned from reinforcement learning are 1) the power of
learning from experience as opposed to labeled training
examples, 2) the central role of modifiable evaluation
functions
in organizing sequential behavior, and 3) that learning and
planning could be radically similar.
Reinforcement Learning and Information
Access (AAAI-SS 3/26/96)
Constructive Induction Needs a Methodology based
on Continuing Learning (ICML94, Workshop on
Constructive Induction, panel remarks)