# scaling-up-brl/scaling-up-brl.github.io

Scaling Up Budgeted Reinforcement Learning
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
assets
code/scaling-up-brl
_config.yml

# Abstract

Can we learn a control policy able to adapt its behaviour in real time so as to take any desired amount of risk? The general Reinforcement Learning framework solely aims at optimising a total reward in expectation, which may not be desirable in critical applications. In stark contrast, the Budgeted Markov Decision Process (BMDP) framework is a formalism in which the notion of risk is implemented as a hard constraint on a failure signal. Existing algorithms solving BMDPs rely on strong assumptions and have so far only been applied to toy-examples. In this work, we relax some of these assumptions and demonstrate the scalability of our approach on two practical problems: a spoken dialogue system and an autonomous driving task. On both examples, we reach similar performances as Lagrangian Relaxation methods with a significant improvement in sample and memory efficiency.

# Risk-Sensitive Exploration

We compare two approaches for contructing a batch of samples. The animations display the trajectories collected in each intermediate subbatch. The first row corresponds to a classical risk neutral epsilon-greedy exploration policy while the second row showcases a risk-sensitive exploration strategy introduced in the paper. Each animation corresponds to a different seed.

## Risk-sensitive strategy

We display the evolution in the budgeted policy behavior with respect to the budget. The policies have been learnt with a risk-sensitive exploration. When the budget is low, the agent takes the safest path on the left. When the budget increases, it gradually switches to the other lane, earning higher rewards but also costs. This gradual process could not be achieved with a deterministic policy as it would chose either one path or the other. Each animation corresponds to a different seed.

# Scalable Budgeted Fitted-Q

## Driving styles

We show samples of driving styles emerging from constraining the time spent on the opposite lane with different constraint budget values. The budget enables to control in real-time the tradeoff between efficiency and safety.

Budget Behaviours

## Dialogue examples

In the following table, we display two dialogues done with the same BFTQ policy. The policy is given two budgets to respect in expectation, and . For budget 0, one can see that the system never uses the ask_num_pad action. Instead, it uses ask_oral, an action subject to recognition errors. The system keeps asking for the same slot 2, because it has the lowest speech recognition score. It eventually summarizes the form to the user, but then reaches the maximum dialogue length and thus faces a dialogue failure. For budget 0.5, the system first asks in a safe way, with ask_oral. It may want to ask_num_pad if one of the speech recognition score is low. Then, the system proceeds to a confirmation of the slot values. If it is incorrect, the system continues the dialogue using unsafe the ask_num_pad action to be certain of the slot values.

turn
turn 0 valid slots : [0, 0, 0]
srs : [ None None None ]
user says "INFORM"
valid slots : [0, 0, 0]
srs : [ None None None ]
user says "INFORM"
turn 1 valid slots : [0, 0, 0]
srs : [ None 0.48 None ]
user says "INFORM"
valid slots : [0, 0, 1]
srs : [ None None 0.56 ]
user says "INFORM"
turn 2 valid slots : [0, 0, 0]
srs : [ None 0.48 0.22 ]
user says "INFORM"
valid slots : [0, 0, 1]
srs : [ 0.30 None 0.56 ]
user says "INFORM"
turn 3 valid slots : [0, 0, 0]
srs : [ 0.62 0.48 0.22 ]
user says "INFORM"
valid slots : [0, 0, 1]
srs : [ 0.30 0.54 0.56 ]
user says "INFORM"
turn 4 valid slots : [0, 0, 0]
srs : [ 0.62 0.48 0.66 ]
user says "INFORM"
valid slots : [0, 0, 1]
srs : [ 0.68 0.54 0.56 ]
user says "INFORM"
turn 5 valid slots : [0, 1, 0]
srs : [ 0.62 0.56 0.66 ]
user says "INFORM"
valid slots : [0, 1, 1]
srs : [ 0.68 1.00 0.56 ]
system says "SUMMARIZE_AND_INFORM"
user says "DENY_SUMMARIZE"
turn 6 valid slots : [0, 1, 0]
srs : [ 0.62 0.56 0.14 ]
user says "INFORM"
valid slots : [0, 1, 1]
srs : [ 0.68 1.00 0.56 ]
user says "INFORM"
turn 7 valid slots : [0, 1, 1]
srs : [ 0.62 0.56 0.30 ]
user says "INFORM"
valid slots : [0, 1, 1]
srs : [ 0.68 1.00 1.00 ]
system says "SUMMARIZE_AND_INFORM"
user says "DENY_SUMMARIZE"
turn 8 valid slots : [0, 1, 1]
srs : [ 0.62 0.56 0.49 ]
user says "INFORM"
valid slots : [0, 1, 1]
srs : [ 0.68 1.00 1.00 ]
user hangs up !
turn 9 valid slots : [0, 1, 1]
srs : [ 0.62 0.56 0.65 ]
system says "SUMMARIZE_AND_INFORM"
max size reached !

# How to reproduce

Install pycairo, numpy, scipy, highway-env and pytorch.

Change python path to the path of this repository:

export PYTHONPATH="scaling-up-brl"

Navigate to budgeted-rl folder :

cd scaling-up-brl/budgeted-rl/

Run main script using any config file. Choose the range of seeds you want to test on:

python main/egreedy/main-egreedy.py config/slot-filling.json 0 6

python main/egreedy/main-egreedy.py config/corridors.json 0 4

python main/egreedy/main-egreedy.py config/highway-easy.json 0 10

## Environment parameters

### Corridors

Parameter Description Value
- Size of the environment 7 x 6
- Standard deviation of the Gaussian noise applied to actions (0.25, 0.25)
H Episode duration 9

### Slot-Filling

Parameter Description Value
ser Sentence Error Rate 0.6
Gaussian mean for misunderstanding -0.25
Gaussian mean for understanding 0.25
Gaussian standard deviation 0.6
Probability of hang-up 0.25
H Episode duration 10
- Number of slots 3

### Highway-Env

Parameter Description Value
Number of vehicles 2 - 6
Standard deviation of vehicles initial positions 100 m
Standard deviation of vehicles initial velocities 3 m/s
H Episode duration 15 s

## Algorithms parameters

### Corridors

Parameters BFTQ(risk-sensitive) BFTQ(risk-neutral)
architecture 256x128x64 256x128x64
regularisation 0.001 0.001
activation relu relu
size beta encoder 3 3
initialisation xavier xavier
loss function L2 L2
learning rate 0.001 0.001
epoch (NN) 1000 5000
normalize reward true true
epoch (FTQ) 12 12
0:0.01:1 -
beta for duplication 0:0.1:1 0:0.1:1
(1,1) (1,1)
5000 5000
10 10
4 4
1000 1000
decay epsilon scheduling 0.001 0.001

### Slot-Filling

Parameters BFTQ FTQ
architecture 256x128x64 128x64x32
regularisation 0.0005 0.0005
activation relu relu
size beta encoder 50 -
initialisation xavier xavier
loss function L2 L2
learning rate 0.001 0.001
epoch (NN) 5000 5000
normalize reward true true
epoch (FTQ) 11 11
0:0.01:1 -
beta for duplication - -
(1,1) (1,1)
5000 5000
10 10
6 6
1000 1000
decay epsilon scheduling 0.001 0.001

### Highway-Env

Parameters BFTQ FTQ
architecture 256x128x64 128x64x32
regularisation 0.0005 0
activation relu relu
size beta encoder 50 -
initialisation xavier xavier
loss function L2 L2