Lionel Blondé will defend on Tuesday the 21st of December 2021 at 15:00 his PhD thesis entitled:
"Counterfactual Interactive Learning: Designing Proactive Artificial Agents that Learn from the Mistakes of other Decision Makers"
Due to the health measures currently in force, the thesis will be defended by video conference publicly, with up to 150 people connected to the same zoom room. In order to facilitate the organisation of the zoom session and following the requirements of the dean's office, if you are interested in attending the defense, please send an email to email@example.com with subject "Defense".
You can check the papers of Lionel here.
The research endeavours laid out in this thesis set out to address weaknesses of reinforcement learning.
As the branch of machine learning that orchestrates a reward-driven decision maker that must learn
from its evolving world via interaction, reinforcement learning relies heavily on said reward’s design.
Modelling a reinforcement signal able to convey the right incentive to the agent for it to carry out
the task at hand is far from obvious, and overall fairly tedious in terms of engineering and tweaking
required. Besides, the interactive nature of the domain puts the agent and its world at satefy risk. The
longer it takes to iterate over and empirically validate reward designs, the less sensible the suffered costs
gradually become. Imitation learning bypasses this time-consuming hurdle by enticing the agent to
mimic an expert instead of trying to maximize a potentially ill-designed reward signal. In effect, the
shift from reinforcement to imitation learning trades the reward feedback the agent receives upon interaction
off for the access to a dataset describing how an expert decision-making policy — assumed
to perform optimally at the task — maps situations to decisions. Depite alleviating the reliance on potentially
satefty-critical reward modelling, state-of-the-art imitation learning agents still suffered from
severe sample-inefficiency in the number of interactions that need be carried out in the environment.
As such, we first set out to address the high sample complexity suffered by Generative Adversarial Imitation
Learning, the state-of-the-art approach to performing imitation (albeit sample-inefficient). To that
end, we build on the strengths of off-policy learning to design a novel adversarial imitation approach. Albeit
coordinating infamously unstable routines — adversarial learning, and off-policy learning — our
novel method remains remarkably stable, and dramatically shrinks the amount of interactions with the
environment necessary to learn well-behaved imitation policies, by up to several orders of magnitude.
In an attempt to decipher and ultimately understand what made the previous method so stable in practice
despite overwhelming odds of instability, we perform an in-depth review, qualitative and quantitative,
of the method. Crucially, we show that forcing the adversarially learned reward function to be local
Lipschitz-continuous is a sine qua non condition for the method to perform well. We then study the effects
of this necessary condition and provide several theoretical results involving the local Lipschitzness of
the state-value function. We complement these guarantees with empirical evidence attesting to the strong
positive effect that the consistent satisfaction of the Lipschitzness constraint on the reward has on imitation
performance. Finally, we tackle a generic pessimistic reward preconditioning add-on spawning a
large class of reward shaping methods, which makes the base method it is plugged into provably more robust,
as shown in several additional theoretical guarantees. We then discuss these through a fine-grained
lens and share our insights. Crucially, the derived guarantees are valid for any reward satisfying the
Lipschitzness condition, nothing is specific to imitation. As such, these may be of independent interest.
While imitation learning seems difficult to outperform once the agent has access to past experiences
from an expert behaving optimally, the performance of the mimicking agent quickly deteriorates as
the expert stray from optimality. Intuitively, it stands to reason that the agent should only imitate another
if the latter solves the task (near-) optimally. If not, then imitation is not a well suited approach.
Datasets of optimal interactive experiences are difficult to come by, and deterringly tedious to design
from scratch. Even then, these rarely perfectly align with the practitioner’s desiderata for the set task.
As a relaxation, interaction traces of sub-optimal behaviors for closely related (yet not exactly aligned)
tasks are far more readily available (e.g. due to continually recording sensors present in robots, cars). As
such, after studying how to only learn from an external source of putatively optimal data coming from
an expert policy, we then study how to only learn from an external source of putatively sub-optimal
data provided through a arbitrary dataset — this last one is the offline reinforcement learning setting.
While in the imitation learning setting we had no reward from the environment, we were still able to
interact with the environment. In the offline reinforcement learning setting, the situation is reversed.
The agent has access to logged interactive experiences from external sources, typically distinct agents.
Said dataset of interactive traces is annotated with rewards perceived by the agents upon interaction.
Yet, the agent learned via offline reinforcement learning is not allowed to interact with the world, and
is therefore unable to witness its own impact on it when trying to make better decisions in it.
The performance of state-of-the-art baselines in the offline reinforcement learning regime varies widely
over the spectrum of dataset qualities — ranging from “far-from-optimal” random data to “close-tooptimal”
expert demonstrations. We re-implement these under a fair, unified, highly factorized, and
open-sourced framework (for which we release the codebase), and show that when a given baseline outperforms
its competing counterparts on one end of the spectrum, it never does on the other end. This consistent
trend prevents us to name a victor that outperforms the rest across the board. We attribute the asymmetry
in performance between the two ends of the quality spectrum to the amount of inductive bias injected into
the agent to entice it to posit that what the behavior underlying the offline dataset is optimal for the task.
The more bias is injected, the higher the agent performs, provided the dataset is close-to-optimal. Otherwise,
its effect is brutally detrimental. Adopting an advantage-weighted regression template as base,
we conduct an investigation which corroborates that injections of such optimality inductive bias, when
not done parsimoniously, makes the agent subpar in the datasets it was dominant as soon as the offline
policy is sub-optimal. To design methods that perform well across the whole spectrum, we revisit the generalized
policy iteration scheme for the offline regime, and study the impact of 9 distinct newly-introduced
proposal distributions over actions, involved in the proposed generalization of the policy evaluation and
policy improvement update rules. We show that certain orchestrations strike the right balance and can
improve the performance on one end of the spectrum without harming it on the other end.
In neither of the two studied settings (a) imitation learning and b) offline reinforcement learning)
does the agent receive direct reward feedback from the world about the quality of the decisions it
makes. It is never told how well it performed, and must infer how well it performed from traces of other
policies, which a priori do not align with what the decision maker would have done. In that light, both
studied settings can be cast as counterfactual learning, in which sequential decision-making agents
must learn how to interact with their world, only from traces that a priori do not coincide with their
own. In fine, we are concerned with decision makers that learn from the choices of others.